Add files using upload-large-folder tool

wandb/run-20260112_222347-jo5wgffb/files/code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py

@@ -0,0 +1,454 @@
+# DQN with Step Penalty & Time Limit Bootstrap for RAGEN Rubik's Cube 2x2
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+import re
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+# 假设 ragen 库在当前目录的上两级，请根据实际情况调整
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.rubikscube.env import RubiksCube2x2Env
+from ragen.env.rubikscube.config import RubiksCube2x2Config
+
+
+class RubiksCubeWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: RubiksCube2x2Env):
+        super().__init__()
+        self._env = env
+        self._colors = ['W', 'O', 'G', 'R', 'B', 'Y']
+        self._color_to_idx = {c: i for i, c in enumerate(self._colors)}
+        self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(24 * len(self._colors),), dtype=np.float32)
+        self.action_space = gym.spaces.Discrete(12)
+        self._face_pat = re.compile(r"\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])\n\s*\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])")
+        
+        # [关键修改] 每步惩罚
+        # 如果20步都没解出来，累计惩罚是 20 * -0.05 = -1.0，抵消掉最后可能获得的 +1.0
+        # 这迫使 agent 寻找更短路径
+        self.step_penalty = -0.1
+
+    def _encode_obs(self, text_obs: str) -> np.ndarray:
+        faces_order = ["Up (U):", "Left (L):", "Front (F):", "Right (R):", "Back (B):", "Down (D):"]
+        lines = text_obs.splitlines()
+        blocks: List[str] = []
+        i = 0
+        while i < len(lines):
+            line = lines[i]
+            for header in faces_order:
+                if line.startswith(header):
+                    content = line[len(header):].strip()
+                    next_line = lines[i + 1] if i + 1 < len(lines) else ""
+                    block = f"{content}\n{next_line}"
+                    blocks.append(block)
+                    break
+            i += 1
+        if len(blocks) != 6:
+            return np.zeros(24 * len(self._colors), dtype=np.float32)
+        stickers: List[int] = []
+        for blk in blocks:
+            m = self._face_pat.search(blk)
+            if not m:
+                return np.zeros(24 * len(self._colors), dtype=np.float32)
+            c0, c1, c2, c3 = m.group(1), m.group(2), m.group(3), m.group(4)
+            stickers.extend([c0, c1, c2, c3])
+        grid = np.zeros((24, len(self._colors)), dtype=np.float32)
+        for idx, ch in enumerate(stickers):
+            cidx = self._color_to_idx.get(ch, None)
+            if cidx is not None:
+                grid[idx, cidx] = 1.0
+        return grid.reshape(-1)
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        mapped = int(action) + 1
+        text_obs, reward, done, info = self._env.step(mapped)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        
+        # [关键修改] 应用惩罚
+        # 此时 reward 包含原本的稀疏奖励 (0 或 1) 加上每步惩罚
+        reward = float(reward) + self.step_penalty
+        
+        return obs, reward, terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "RubiksCube2x2DQN"
+    total_timesteps: int = 400_000
+    learning_rate: float = 2.5e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 200_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 4
+    learning_starts: int = 5000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Rubik specific
+    scramble_depth: int = 2 # 初始难度
+    max_steps_env: int = 20 # 硬性时间限制
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, scramble_depth, max_steps_env, capture_video=False):
+    def thunk():
+        # [关键修改] 内部 max_steps 设为 1000，防止 env 内部发出 done=True
+        # 我们完全依赖外部 TimeLimit wrapper 来处理超时
+        config = RubiksCube2x2Config(scramble_depth=scramble_depth, max_steps=20, render_mode='text')
+        env = RubiksCube2x2Env(config)
+        env = RubiksCubeWrapper(env)
+        # 外部限制设为 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps_env)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_shape: Tuple[int, ...]):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.scramble_depth, args.max_steps_env, args.capture_video)()
+    obs_shape = env.observation_space.shape
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_shape)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    # 评估函数
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                state, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                traj_states = [np.asarray(state).tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.max_steps_env)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        action = int(torch.argmax(q, dim=1).item())
+                    next_state, reward, terminated, truncated, info = env_eval.step(action)
+                    
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_state
+                    traj_states.append(np.asarray(state).tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    # 主训练循环
+    global_step = 0
+    start_time = time.time()
+
+    obs, _ = env.reset(seed=args.seed)
+    ep_return = 0.0
+    ep_len = 0
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        if np.random.rand() < epsilon or global_step < args.learning_starts:
+            action = env.action_space.sample()
+        else:
+            with torch.no_grad():
+                q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+                action = int(torch.argmax(q_values, dim=1).item())
+
+        next_obs, reward, terminated, truncated, info = env.step(action)
+        
+        # 判断是否需要 Reset (任何结束都 Reset)
+        real_done = bool(terminated) or bool(truncated)
+
+        # [核心逻辑]
+        # 1. 即使是 truncated (超时)，done 也记为 False，以便进行 bootstrap (计算未来价值)
+        # 2. 只有 terminated (真正解开了)，done 才记为 True
+        # 3. Step Penalty 已经包含在 reward 中，DQN 会学到"为了避免扣分，必须在未来几步内解决"
+        rb.add(
+            obs.astype(np.float32), 
+            int(action), 
+            float(reward), 
+            bool(terminated), # 注意：只用 terminated
+            next_obs.astype(np.float32)
+        )
+
+        obs = next_obs
+        ep_return += float(reward)
+        ep_len += 1
+        global_step += 1
+
+        # 训练过程
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        # 结束时 Reset
+        if real_done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/episodic_return": float(ep_return),
+                        "train/episodic_length": int(ep_len),
+                        "train/success": float(1.0 if succ else 0.0),
+                        "train/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) >= 1 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs, _ = env.reset()
+            ep_return, ep_len = 0.0, 0
+
+        # 更新目标网络
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        # 打印日志
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        # 评估
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    # 评估环境也同样设置：内部1000，外部20
+                    return make_env(0, run_name, args.seed + 9999, args.scramble_depth, args.max_steps_env, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_223423-ozob7eqw/files/code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py

@@ -0,0 +1,454 @@
+# DQN with Step Penalty & Time Limit Bootstrap for RAGEN Rubik's Cube 2x2
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+import re
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+# 假设 ragen 库在当前目录的上两级，请根据实际情况调整
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.rubikscube.env import RubiksCube2x2Env
+from ragen.env.rubikscube.config import RubiksCube2x2Config
+
+
+class RubiksCubeWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: RubiksCube2x2Env):
+        super().__init__()
+        self._env = env
+        self._colors = ['W', 'O', 'G', 'R', 'B', 'Y']
+        self._color_to_idx = {c: i for i, c in enumerate(self._colors)}
+        self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(24 * len(self._colors),), dtype=np.float32)
+        self.action_space = gym.spaces.Discrete(12)
+        self._face_pat = re.compile(r"\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])\n\s*\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])")
+        
+        # [关键修改] 每步惩罚
+        # 如果20步都没解出来，累计惩罚是 20 * -0.05 = -1.0，抵消掉最后可能获得的 +1.0
+        # 这迫使 agent 寻找更短路径
+        self.step_penalty = -0.1
+
+    def _encode_obs(self, text_obs: str) -> np.ndarray:
+        faces_order = ["Up (U):", "Left (L):", "Front (F):", "Right (R):", "Back (B):", "Down (D):"]
+        lines = text_obs.splitlines()
+        blocks: List[str] = []
+        i = 0
+        while i < len(lines):
+            line = lines[i]
+            for header in faces_order:
+                if line.startswith(header):
+                    content = line[len(header):].strip()
+                    next_line = lines[i + 1] if i + 1 < len(lines) else ""
+                    block = f"{content}\n{next_line}"
+                    blocks.append(block)
+                    break
+            i += 1
+        if len(blocks) != 6:
+            return np.zeros(24 * len(self._colors), dtype=np.float32)
+        stickers: List[int] = []
+        for blk in blocks:
+            m = self._face_pat.search(blk)
+            if not m:
+                return np.zeros(24 * len(self._colors), dtype=np.float32)
+            c0, c1, c2, c3 = m.group(1), m.group(2), m.group(3), m.group(4)
+            stickers.extend([c0, c1, c2, c3])
+        grid = np.zeros((24, len(self._colors)), dtype=np.float32)
+        for idx, ch in enumerate(stickers):
+            cidx = self._color_to_idx.get(ch, None)
+            if cidx is not None:
+                grid[idx, cidx] = 1.0
+        return grid.reshape(-1)
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        mapped = int(action) + 1
+        text_obs, reward, done, info = self._env.step(mapped)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        
+        # [关键修改] 应用惩罚
+        # 此时 reward 包含原本的稀疏奖励 (0 或 1) 加上每步惩罚
+        reward = float(reward) + self.step_penalty
+        
+        return obs, reward, terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "RubiksCube2x2DQN"
+    total_timesteps: int = 400_000
+    learning_rate: float = 2.5e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 200_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 4
+    learning_starts: int = 5000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Rubik specific
+    scramble_depth: int = 3 # 初始难度
+    max_steps_env: int = 20 # 硬性时间限制
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, scramble_depth, max_steps_env, capture_video=False):
+    def thunk():
+        # [关键修改] 内部 max_steps 设为 1000，防止 env 内部发出 done=True
+        # 我们完全依赖外部 TimeLimit wrapper 来处理超时
+        config = RubiksCube2x2Config(scramble_depth=scramble_depth, max_steps=20, render_mode='text')
+        env = RubiksCube2x2Env(config)
+        env = RubiksCubeWrapper(env)
+        # 外部限制设为 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps_env)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_shape: Tuple[int, ...]):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.scramble_depth, args.max_steps_env, args.capture_video)()
+    obs_shape = env.observation_space.shape
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_shape)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    # 评估函数
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                state, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                traj_states = [np.asarray(state).tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.max_steps_env)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        action = int(torch.argmax(q, dim=1).item())
+                    next_state, reward, terminated, truncated, info = env_eval.step(action)
+                    
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_state
+                    traj_states.append(np.asarray(state).tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    # 主训练循环
+    global_step = 0
+    start_time = time.time()
+
+    obs, _ = env.reset(seed=args.seed)
+    ep_return = 0.0
+    ep_len = 0
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        if np.random.rand() < epsilon or global_step < args.learning_starts:
+            action = env.action_space.sample()
+        else:
+            with torch.no_grad():
+                q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+                action = int(torch.argmax(q_values, dim=1).item())
+
+        next_obs, reward, terminated, truncated, info = env.step(action)
+        
+        # 判断是否需要 Reset (任何结束都 Reset)
+        real_done = bool(terminated) or bool(truncated)
+
+        # [核心逻辑]
+        # 1. 即使是 truncated (超时)，done 也记为 False，以便进行 bootstrap (计算未来价值)
+        # 2. 只有 terminated (真正解开了)，done 才记为 True
+        # 3. Step Penalty 已经包含在 reward 中，DQN 会学到"为了避免扣分，必须在未来几步内解决"
+        rb.add(
+            obs.astype(np.float32), 
+            int(action), 
+            float(reward), 
+            bool(terminated), # 注意：只用 terminated
+            next_obs.astype(np.float32)
+        )
+
+        obs = next_obs
+        ep_return += float(reward)
+        ep_len += 1
+        global_step += 1
+
+        # 训练过程
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        # 结束时 Reset
+        if real_done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/episodic_return": float(ep_return),
+                        "train/episodic_length": int(ep_len),
+                        "train/success": float(1.0 if succ else 0.0),
+                        "train/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) >= 1 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs, _ = env.reset()
+            ep_return, ep_len = 0.0, 0
+
+        # 更新目标网络
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        # 打印日志
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        # 评估
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    # 评估环境也同样设置：内部1000，外部20
+                    return make_env(0, run_name, args.seed + 9999, args.scramble_depth, args.max_steps_env, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_223423-ozob7eqw/files/config.yaml

@@ -0,0 +1,167 @@
+_wandb:
+    value:
+        cli_version: 0.23.0
+        code_path: code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+        e:
+            1qvdn1dfs9lrx9hpw1cc1d9q3gh3ht3r:
+                args:
+                    - --track
+                codePath: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                codePathLocal: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "2576980377600"
+                        used: "3724288000"
+                email: haoyu-wa22@mails.tsinghua.edu.cn
+                executable: /root/local/miniconda3/envs/ragen/bin/python
+                git:
+                    commit: 8d73639b99b38265453f898b8d6af7d4af50d56e
+                    remote: https://github.com/mll-lab-nu/RAGEN.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-cff454b9-f118-535b-5e1f-6d503e78e52e
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9e77c178-188c-f869-780e-858c20a7d322
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7
+                host: pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0
+                memory:
+                    total: "2163639439360"
+                os: Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35
+                program: /mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                python: CPython 3.10.19
+                root: /mnt/general/wanghy/RAGEN
+                startedAt: "2026-01-12T14:34:23.071083Z"
+                writerId: 1qvdn1dfs9lrx9hpw1cc1d9q3gh3ht3r
+        m:
+            - "1": global_step
+              "6":
+                - 3
+              "7": []
+            - "2": train/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": rollout/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": eval/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": losses/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": charts/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": perf/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+        python_version: 3.10.19
+        t:
+            "1":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "2":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "3":
+                - 7
+                - 13
+                - 16
+                - 61
+            "4": 3.10.19
+            "5": 0.23.0
+            "6": 4.57.3
+            "12": 0.23.0
+            "13": linux-x86_64
+batch_size:
+    value: 128
+buffer_size:
+    value: 200000
+capture_video:
+    value: false
+cuda:
+    value: true
+end_e:
+    value: 0.05
+env_id:
+    value: RubiksCube2x2DQN
+eval_episodes:
+    value: 4000
+eval_splits:
+    value: 2
+exp_name:
+    value: dqn_rubikscube
+exploration_fraction:
+    value: 0.2
+gamma:
+    value: 0.99
+hidden_size:
+    value: 256
+learning_rate:
+    value: 0.00025
+learning_starts:
+    value: 5000
+max_steps_env:
+    value: 20
+scramble_depth:
+    value: 3
+seed:
+    value: 1
+start_e:
+    value: 1
+target_network_frequency:
+    value: 4000
+torch_deterministic:
+    value: true
+total_timesteps:
+    value: 400000
+track:
+    value: true
+train_frequency:
+    value: 4
+wandb_entity:
+    value: null
+wandb_project_name:
+    value: cleanRL

wandb/run-20260112_223423-ozob7eqw/files/diff.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_223423-ozob7eqw/files/diff_8d73639b99b38265453f898b8d6af7d4af50d56e.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_223423-ozob7eqw/files/requirements.txt

@@ -0,0 +1,313 @@
+colorama==0.4.6
+psutil==7.1.3
+ragen==0.1
+Farama-Notifications==0.0.4
+MarkupSafe==2.1.5
+PyJWT==2.10.1
+PySocks==1.7.1
+absl-py==2.3.1
+accelerate==1.12.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.2
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+airportsdata==20250909
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anthropic==0.75.0
+antlr4-python3-runtime==4.9.3
+anyio==4.12.0
+astor==0.8.1
+asttokens==3.0.1
+async-timeout==5.0.1
+attrs==25.4.0
+beautifulsoup4==4.14.3
+black==25.11.0
+blake3==1.0.8
+blinker==1.9.0
+blis==1.3.3
+cachetools==6.2.2
+catalogue==2.0.10
+certifi==2025.11.12
+cffi==2.0.0
+charset-normalizer==3.4.4
+cleantext==1.1.4
+click==8.2.1
+cloudpathlib==0.23.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+coloredlogs==15.0.1
+colorful==0.5.8
+compressed-tensors==0.9.3
+confection==0.1.5
+contourpy==1.3.2
+cryptography==46.0.3
+cupy-cuda12x==13.6.0
+cycler==0.12.1
+cymem==2.0.13
+Cython==3.2.2
+datasets==4.4.1
+debugpy==1.8.17
+decorator==5.2.1
+Deprecated==1.3.1
+depyf==0.18.0
+dill==0.4.0
+diskcache==5.6.3
+distlib==0.4.0
+distro==1.9.0
+dnspython==2.8.0
+docstring_parser==0.17.0
+einops==0.8.1
+email-validator==2.3.0
+eval_type_backport==0.2.2
+exceptiongroup==1.3.1
+executing==2.2.1
+faiss-gpu==1.7.2
+fastapi==0.123.1
+fastapi-cli==0.0.16
+fastapi-cloud-cli==0.5.2
+fastar==0.8.0
+fastrlock==0.8.3
+filelock==3.19.1
+flash_attn==2.8.3
+Flask==3.1.2
+flatbuffers==25.9.23
+fonttools==4.61.0
+frozenlist==1.8.0
+fsspec==2025.9.0
+gdown==5.2.0
+gguf==0.17.1
+gitdb==4.0.12
+GitPython==3.1.45
+google-api-core==2.28.1
+google-auth==2.43.0
+googleapis-common-protos==1.72.0
+grpcio==1.76.0
+gym==0.26.2
+gym-notices==0.1.0
+gym-sokoban==0.0.6
+gymnasium==1.2.2
+h11==0.16.0
+hf-xet==1.2.0
+html2text==2025.4.15
+httpcore==1.0.9
+httptools==0.7.1
+httpx==0.28.1
+httpx-sse==0.4.3
+huggingface-hub==0.36.0
+humanfriendly==10.0
+hydra-core==1.3.2
+idna==3.11
+ImageIO==2.37.2
+importlib_metadata==8.0.0
+interegular==0.3.3
+ipython==8.37.0
+itsdangerous==2.2.0
+jedi==0.19.2
+Jinja2==3.1.6
+jiter==0.12.0
+joblib==1.5.2
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+kimina-client==0.2.1
+kiwisolver==1.4.9
+lark==1.2.2
+llguidance==0.7.30
+llvmlite==0.44.0
+lm-format-enforcer==0.10.12
+loguru==0.7.3
+Markdown==3.10
+markdown-it-py==4.0.0
+matplotlib==3.10.7
+matplotlib-inline==0.2.1
+mcp==1.22.0
+mdurl==0.1.2
+mistral_common==1.8.6
+mpmath==1.3.0
+msgpack==1.1.2
+msgspec==0.20.0
+multidict==6.7.0
+multiprocess==0.70.18
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+nest-asyncio==1.6.0
+networkx==3.3
+ninja==1.13.0
+nltk==3.9.2
+numba==0.61.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-cusparselt-cu12==0.6.2
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+omegaconf==2.3.0
+onnxruntime==1.23.2
+openai==2.8.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opencv-python-headless==4.12.0.88
+opentelemetry-api==1.38.0
+opentelemetry-exporter-otlp==1.26.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+opentelemetry-exporter-prometheus==0.59b0
+opentelemetry-proto==1.26.0
+opentelemetry-sdk==1.38.0
+opentelemetry-semantic-conventions==0.59b0
+opentelemetry-semantic-conventions-ai==0.4.13
+orjson==3.11.4
+outlines==0.1.11
+outlines_core==0.1.26
+packaging==25.0
+pandas==2.3.3
+parso==0.8.5
+partial-json-parser==0.2.1.1.post7
+pathspec==0.12.1
+peft==0.18.0
+pexpect==4.9.0
+pillow==11.3.0
+pip==25.3
+platformdirs==4.5.0
+preshed==3.0.12
+prometheus_client==0.23.1
+prometheus-fastapi-instrumentator==7.1.0
+prompt_toolkit==3.0.52
+propcache==0.4.1
+proto-plus==1.26.1
+protobuf==4.25.8
+psutil==7.1.3
+ptyprocess==0.7.0
+pure_eval==0.2.3
+py-cpuinfo==9.0.0
+py-spy==0.4.1
+pyarrow==22.0.0
+pyasn1==0.6.1
+pyasn1_modules==0.4.2
+pybind11==3.0.1
+pycountry==24.6.1
+pycparser==2.23
+pydantic==2.12.5
+pydantic_core==2.41.5
+pydantic-extra-types==2.10.6
+pydantic-settings==2.12.0
+pygame==2.6.1
+Pygments==2.19.2
+pyjnius==1.7.0
+pylatexenc==2.10
+pyparsing==3.2.5
+pyserini==1.2.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.1
+python-json-logger==4.0.0
+python-multipart==0.0.20
+pytokens==0.3.0
+pytz==2025.2
+PyYAML==6.0.3
+pyzmq==27.1.0
+ragen==0.1
+rank-bm25==0.2.2
+RapidFuzz==3.14.3
+ray==2.52.1
+referencing==0.37.0
+regex==2025.11.3
+requests==2.32.5
+rich==14.2.0
+rich-toolkit==0.17.0
+rignore==0.7.6
+rpds-py==0.30.0
+rsa==4.9.1
+safetensors==0.7.0
+scikit-learn==1.7.2
+scipy==1.15.3
+sentencepiece==0.2.1
+sentry-sdk==2.46.0
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.17.0
+smart_open==7.5.0
+smmap==5.0.2
+sniffio==1.3.1
+soupsieve==2.8
+spacy==3.8.11
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.2
+sse-starlette==3.0.3
+stack-data==0.6.3
+starlette==0.50.0
+sympy==1.13.1
+tabulate==0.9.0
+tenacity==9.1.2
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+tensordict==0.8.3
+thefuzz==0.22.1
+thinc==8.3.10
+threadpoolctl==3.6.0
+tiktoken==0.12.0
+together==1.5.31
+tokenizers==0.22.1
+tomli==2.3.0
+torch==2.6.0+cu124
+torchaudio==2.6.0
+torchdata==0.11.0
+torchvision==0.21.0
+tqdm==4.67.1
+traitlets==5.14.3
+transformers==4.57.3
+triton==3.2.0
+typer==0.19.2
+typer-slim==0.20.0
+typing_extensions==4.15.0
+typing-inspection==0.4.2
+tzdata==2025.2
+urllib3==2.5.0
+uvicorn==0.38.0
+uvloop==0.22.1
+verl==0.5.0.dev0
+virtualenv==20.35.4
+vllm==0.8.5
+wandb==0.23.0
+wasabi==1.1.3
+watchfiles==1.1.1
+wcwidth==0.2.14
+weasel==0.4.3
+websockets==15.0.1
+Werkzeug==3.1.4
+wheel==0.45.1
+wrapt==2.0.1
+xformers==0.0.29.post2
+xgrammar==0.1.18
+xxhash==3.6.0
+yarl==1.22.0
+zipp==3.23.0
+typeguard==4.4.4
+tyro==1.0.3
+verl==0.5.0.dev0
+autocommand==2.2.2
+backports.tarfile==1.2.0
+importlib_metadata==8.0.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+jaraco.context==5.3.0
+jaraco.functools==4.0.1
+jaraco.text==3.12.1
+more-itertools==10.3.0
+packaging==24.2
+platformdirs==4.2.2
+tomli==2.0.1
+typeguard==4.3.0
+typing_extensions==4.12.2
+wheel==0.45.1
+zipp==3.19.2

wandb/run-20260112_223423-ozob7eqw/files/wandb-metadata.json

@@ -0,0 +1,64 @@
+{
+  "os": "Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.19",
+  "startedAt": "2026-01-12T14:34:23.071083Z",
+  "args": [
+    "--track"
+  ],
+  "program": "/mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePath": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePathLocal": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "git": {
+    "remote": "https://github.com/mll-lab-nu/RAGEN.git",
+    "commit": "8d73639b99b38265453f898b8d6af7d4af50d56e"
+  },
+  "email": "haoyu-wa22@mails.tsinghua.edu.cn",
+  "root": "/mnt/general/wanghy/RAGEN",
+  "host": "pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0",
+  "executable": "/root/local/miniconda3/envs/ragen/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "2576980377600",
+      "used": "3724288000"
+    }
+  },
+  "memory": {
+    "total": "2163639439360"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-cff454b9-f118-535b-5e1f-6d503e78e52e"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9e77c178-188c-f869-780e-858c20a7d322"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "1qvdn1dfs9lrx9hpw1cc1d9q3gh3ht3r"
+}

wandb/run-20260112_223423-ozob7eqw/files/wandb-summary.json

@@ -0,0 +1 @@
+{"train/loss":0.00112406094558537,"_step":173979,"charts/epsilon":0.05,"eval/episodes":4000,"train/episodic_length":3,"perf/SPS":689,"eval/std_return":0.6024724967788987,"_timestamp":1.7682287212650092e+09,"train/success_rate_100":0.98,"train/success":1,"train/learning_rate":0.00025,"eval/success_rate":0.04625,"_wandb":{"runtime":256},"eval/avg_return":-1.8673750000000005,"train/episodic_return":0.7,"global_step":173979,"_runtime":256.543296735}

wandb/run-20260112_223643-sgt786ss/files/code/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py

@@ -0,0 +1,485 @@
+# DQN with Action Masking for RAGEN Sudoku (4x4, max_step=20)
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.sudoku.env import SudokuEnv
+from ragen.env.sudoku.config import SudokuEnvConfig
+
+
+class SudokuWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: SudokuEnv, grid_size: int):
+        super().__init__()
+        self._env = env
+        self._size = grid_size
+        self._val_dim = self._size + 1
+        self._act_n = self._size * self._size * self._size
+        self.action_space = gym.spaces.Discrete(self._act_n)
+        self.observation_space = gym.spaces.Dict({
+            "observation": gym.spaces.Box(low=0.0, high=1.0, shape=(self._size * self._size * self._val_dim,), dtype=np.float32),
+            "action_mask": gym.spaces.Box(low=0.0, high=1.0, shape=(self._act_n,), dtype=np.float32)
+        })
+
+    def _encode_obs(self, text_obs: str) -> Dict[str, np.ndarray]:
+        vals: List[int] = []
+        for line in text_obs.splitlines():
+            ls = line.strip()
+            if len(ls) == 0:
+                continue
+            if set(ls) <= {'-'}:
+                continue
+            tokens = [t for t in ls.split() if t != '|']
+            if len(tokens) == 0:
+                continue
+            for t in tokens:
+                if t == '.':
+                    vals.append(0)
+                else:
+                    try:
+                        v = int(t)
+                    except ValueError:
+                        v = 0
+                    vals.append(v)
+        target = self._size * self._size
+        if len(vals) < target:
+            vals.extend([0] * (target - len(vals)))
+        if len(vals) > target:
+            vals = vals[:target]
+        grid = np.zeros((target, self._val_dim), dtype=np.float32)
+        mask = np.zeros(self._act_n, dtype=np.float32)
+        for i, v in enumerate(vals):
+            v_clamped = int(v)
+            if v_clamped < 0 or v_clamped > self._size:
+                v_clamped = 0
+            grid[i, v_clamped] = 1.0
+            if v_clamped != 0:
+                start_idx = i * self._size
+                end_idx = start_idx + self._size
+                mask[start_idx:end_idx] = 0.0
+        cur = np.array(vals, dtype=np.int64).reshape(self._size, self._size)
+        box = int(np.sqrt(self._size))
+        for i, v in enumerate(vals):
+            if int(v) == 0:
+                r = i // self._size
+                c = i % self._size
+                row_vals = set(cur[r, :].tolist())
+                col_vals = set(cur[:, c].tolist())
+                br = (r // box) * box
+                bc = (c // box) * box
+                box_vals = set(cur[br:br + box, bc:bc + box].reshape(-1).tolist())
+                start_idx = i * self._size
+                for num in range(1, self._size + 1):
+                    if (num not in row_vals) and (num not in col_vals) and (num not in box_vals):
+                        mask[start_idx + (num - 1)] = 1.0
+        if mask.sum() == 0:
+            for i, v in enumerate(vals):
+                if int(v) == 0:
+                    start_idx = i * self._size
+                    end_idx = start_idx + self._size
+                    mask[start_idx:end_idx] = 1.0
+            if mask.sum() == 0:
+                mask[:] = 1.0
+        return {
+            "observation": grid.reshape(-1),
+            "action_mask": mask,
+        }
+
+    @staticmethod
+    def _decode_action(action_id: int, grid_size: int) -> Tuple[int, int, int]:
+        g = grid_size
+        row = action_id // (g * g)
+        rem = action_id % (g * g)
+        col = rem // g
+        num = (rem % g) + 1
+        return row, col, num
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        row, col, num = self._decode_action(int(action), self._size)
+        act_str = f"{row + 1},{col + 1},{num}"
+        text_obs, reward, done, info = self._env.step(act_str)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        return obs, float(reward), terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "SudokuDQN"
+    total_timesteps: int = 2_000_000
+    learning_rate: float = 3e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 400_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 1
+    learning_starts: int = 10000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Sudoku specific
+    grid_size: int = 4
+    difficulty: str = "easy"
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, grid_size, difficulty, capture_video=False):
+    def thunk():
+        config = SudokuEnvConfig(
+            grid_size=grid_size,
+            difficulty=difficulty,
+            render_mode='text',
+            render_format='simple',
+        )
+        env = SudokuEnv(config)
+        env = SudokuWrapper(env, grid_size)
+        max_steps = 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_dim: int):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity, obs_dim), dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity, obs_dim), dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.grid_size, args.difficulty, args.capture_video)()
+    # Flattened grid only from Dict observation
+    sample_obs, _ = env.reset(seed=args.seed)
+    obs_dim = int(np.prod(sample_obs["observation"].shape))
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(obs_dim, act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(obs_dim, act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_dim)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                obs_dict, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                state = obs_dict['observation']
+                mask = obs_dict['action_mask']
+                traj_states = [state.tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.grid_size * args.grid_size * 6)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        mask_t = torch.tensor(mask, dtype=torch.float32, device=device).unsqueeze(0)
+                        masked_q = q + (mask_t - 1.0) * 1e8
+                        action = int(torch.argmax(masked_q, dim=1).item())
+                    next_obs_dict, reward, terminated, truncated, info = env_eval.step(action)
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_obs_dict['observation']
+                    mask = next_obs_dict['action_mask']
+                    traj_states.append(state.tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    global_step = 0
+    start_time = time.time()
+
+    obs_dict, _ = env.reset(seed=args.seed)
+    obs = obs_dict['observation']
+    mask = obs_dict['action_mask']
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        with torch.no_grad():
+            q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+            mask_t = torch.tensor(mask, dtype=torch.float32, device=device).unsqueeze(0)
+            masked_q = q_values + (mask_t - 1.0) * 1e8
+            greedy_action = int(torch.argmax(masked_q, dim=1).item())
+        if (np.random.rand() < epsilon) or (global_step < args.learning_starts):
+            valid = np.where(mask > 0.5)[0]
+            if len(valid) > 0:
+                action = int(np.random.choice(valid))
+            else:
+                action = int(np.random.randint(0, act_dim))
+        else:
+            action = greedy_action
+
+        next_obs_dict, reward, terminated, truncated, info = env.step(action)
+        done = bool(terminated) or bool(truncated)
+
+        next_obs = next_obs_dict['observation']
+        next_mask = next_obs_dict['action_mask']
+
+        rb.add(obs.astype(np.float32), int(action), float(reward), bool(done), next_obs.astype(np.float32))
+
+        obs = next_obs
+        mask = next_mask
+        global_step += 1
+
+        if done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "rollout/success": float(1.0 if succ else 0.0),
+                        "rollout/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs_dict, _ = env.reset()
+            obs = obs_dict['observation']
+            mask = obs_dict['action_mask']
+
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    return make_env(0, run_name, args.seed + 9999, args.grid_size, args.difficulty, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_223643-sgt786ss/files/config.yaml

@@ -0,0 +1,167 @@
+_wandb:
+    value:
+        cli_version: 0.23.0
+        code_path: code/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+        e:
+            shc54txtsp3x3q5ajbrzfnd1sznlb05g:
+                args:
+                    - --track
+                codePath: cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+                codePathLocal: cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "2576980377600"
+                        used: "3724296192"
+                email: haoyu-wa22@mails.tsinghua.edu.cn
+                executable: /root/local/miniconda3/envs/ragen/bin/python
+                git:
+                    commit: 8d73639b99b38265453f898b8d6af7d4af50d56e
+                    remote: https://github.com/mll-lab-nu/RAGEN.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-cff454b9-f118-535b-5e1f-6d503e78e52e
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9e77c178-188c-f869-780e-858c20a7d322
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7
+                host: pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0
+                memory:
+                    total: "2163639439360"
+                os: Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35
+                program: /mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+                python: CPython 3.10.19
+                root: /mnt/general/wanghy/RAGEN
+                startedAt: "2026-01-12T14:36:43.218427Z"
+                writerId: shc54txtsp3x3q5ajbrzfnd1sznlb05g
+        m:
+            - "1": global_step
+              "6":
+                - 3
+              "7": []
+            - "2": eval/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": losses/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": charts/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": perf/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": train/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": rollout/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+        python_version: 3.10.19
+        t:
+            "1":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "2":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "3":
+                - 7
+                - 13
+                - 16
+                - 61
+            "4": 3.10.19
+            "5": 0.23.0
+            "6": 4.57.3
+            "12": 0.23.0
+            "13": linux-x86_64
+batch_size:
+    value: 128
+buffer_size:
+    value: 400000
+capture_video:
+    value: false
+cuda:
+    value: true
+difficulty:
+    value: easy
+end_e:
+    value: 0.05
+env_id:
+    value: SudokuDQN
+eval_episodes:
+    value: 4000
+eval_splits:
+    value: 2
+exp_name:
+    value: dqn_sudoku
+exploration_fraction:
+    value: 0.2
+gamma:
+    value: 0.99
+grid_size:
+    value: 4
+hidden_size:
+    value: 256
+learning_rate:
+    value: 0.0003
+learning_starts:
+    value: 10000
+seed:
+    value: 1
+start_e:
+    value: 1
+target_network_frequency:
+    value: 4000
+torch_deterministic:
+    value: true
+total_timesteps:
+    value: 2000000
+track:
+    value: true
+train_frequency:
+    value: 1
+wandb_entity:
+    value: null
+wandb_project_name:
+    value: cleanRL

wandb/run-20260112_223643-sgt786ss/files/diff.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_223643-sgt786ss/files/diff_8d73639b99b38265453f898b8d6af7d4af50d56e.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_223643-sgt786ss/files/requirements.txt

@@ -0,0 +1,313 @@
+colorama==0.4.6
+psutil==7.1.3
+ragen==0.1
+Farama-Notifications==0.0.4
+MarkupSafe==2.1.5
+PyJWT==2.10.1
+PySocks==1.7.1
+absl-py==2.3.1
+accelerate==1.12.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.2
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+airportsdata==20250909
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anthropic==0.75.0
+antlr4-python3-runtime==4.9.3
+anyio==4.12.0
+astor==0.8.1
+asttokens==3.0.1
+async-timeout==5.0.1
+attrs==25.4.0
+beautifulsoup4==4.14.3
+black==25.11.0
+blake3==1.0.8
+blinker==1.9.0
+blis==1.3.3
+cachetools==6.2.2
+catalogue==2.0.10
+certifi==2025.11.12
+cffi==2.0.0
+charset-normalizer==3.4.4
+cleantext==1.1.4
+click==8.2.1
+cloudpathlib==0.23.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+coloredlogs==15.0.1
+colorful==0.5.8
+compressed-tensors==0.9.3
+confection==0.1.5
+contourpy==1.3.2
+cryptography==46.0.3
+cupy-cuda12x==13.6.0
+cycler==0.12.1
+cymem==2.0.13
+Cython==3.2.2
+datasets==4.4.1
+debugpy==1.8.17
+decorator==5.2.1
+Deprecated==1.3.1
+depyf==0.18.0
+dill==0.4.0
+diskcache==5.6.3
+distlib==0.4.0
+distro==1.9.0
+dnspython==2.8.0
+docstring_parser==0.17.0
+einops==0.8.1
+email-validator==2.3.0
+eval_type_backport==0.2.2
+exceptiongroup==1.3.1
+executing==2.2.1
+faiss-gpu==1.7.2
+fastapi==0.123.1
+fastapi-cli==0.0.16
+fastapi-cloud-cli==0.5.2
+fastar==0.8.0
+fastrlock==0.8.3
+filelock==3.19.1
+flash_attn==2.8.3
+Flask==3.1.2
+flatbuffers==25.9.23
+fonttools==4.61.0
+frozenlist==1.8.0
+fsspec==2025.9.0
+gdown==5.2.0
+gguf==0.17.1
+gitdb==4.0.12
+GitPython==3.1.45
+google-api-core==2.28.1
+google-auth==2.43.0
+googleapis-common-protos==1.72.0
+grpcio==1.76.0
+gym==0.26.2
+gym-notices==0.1.0
+gym-sokoban==0.0.6
+gymnasium==1.2.2
+h11==0.16.0
+hf-xet==1.2.0
+html2text==2025.4.15
+httpcore==1.0.9
+httptools==0.7.1
+httpx==0.28.1
+httpx-sse==0.4.3
+huggingface-hub==0.36.0
+humanfriendly==10.0
+hydra-core==1.3.2
+idna==3.11
+ImageIO==2.37.2
+importlib_metadata==8.0.0
+interegular==0.3.3
+ipython==8.37.0
+itsdangerous==2.2.0
+jedi==0.19.2
+Jinja2==3.1.6
+jiter==0.12.0
+joblib==1.5.2
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+kimina-client==0.2.1
+kiwisolver==1.4.9
+lark==1.2.2
+llguidance==0.7.30
+llvmlite==0.44.0
+lm-format-enforcer==0.10.12
+loguru==0.7.3
+Markdown==3.10
+markdown-it-py==4.0.0
+matplotlib==3.10.7
+matplotlib-inline==0.2.1
+mcp==1.22.0
+mdurl==0.1.2
+mistral_common==1.8.6
+mpmath==1.3.0
+msgpack==1.1.2
+msgspec==0.20.0
+multidict==6.7.0
+multiprocess==0.70.18
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+nest-asyncio==1.6.0
+networkx==3.3
+ninja==1.13.0
+nltk==3.9.2
+numba==0.61.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-cusparselt-cu12==0.6.2
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+omegaconf==2.3.0
+onnxruntime==1.23.2
+openai==2.8.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opencv-python-headless==4.12.0.88
+opentelemetry-api==1.38.0
+opentelemetry-exporter-otlp==1.26.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+opentelemetry-exporter-prometheus==0.59b0
+opentelemetry-proto==1.26.0
+opentelemetry-sdk==1.38.0
+opentelemetry-semantic-conventions==0.59b0
+opentelemetry-semantic-conventions-ai==0.4.13
+orjson==3.11.4
+outlines==0.1.11
+outlines_core==0.1.26
+packaging==25.0
+pandas==2.3.3
+parso==0.8.5
+partial-json-parser==0.2.1.1.post7
+pathspec==0.12.1
+peft==0.18.0
+pexpect==4.9.0
+pillow==11.3.0
+pip==25.3
+platformdirs==4.5.0
+preshed==3.0.12
+prometheus_client==0.23.1
+prometheus-fastapi-instrumentator==7.1.0
+prompt_toolkit==3.0.52
+propcache==0.4.1
+proto-plus==1.26.1
+protobuf==4.25.8
+psutil==7.1.3
+ptyprocess==0.7.0
+pure_eval==0.2.3
+py-cpuinfo==9.0.0
+py-spy==0.4.1
+pyarrow==22.0.0
+pyasn1==0.6.1
+pyasn1_modules==0.4.2
+pybind11==3.0.1
+pycountry==24.6.1
+pycparser==2.23
+pydantic==2.12.5
+pydantic_core==2.41.5
+pydantic-extra-types==2.10.6
+pydantic-settings==2.12.0
+pygame==2.6.1
+Pygments==2.19.2
+pyjnius==1.7.0
+pylatexenc==2.10
+pyparsing==3.2.5
+pyserini==1.2.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.1
+python-json-logger==4.0.0
+python-multipart==0.0.20
+pytokens==0.3.0
+pytz==2025.2
+PyYAML==6.0.3
+pyzmq==27.1.0
+ragen==0.1
+rank-bm25==0.2.2
+RapidFuzz==3.14.3
+ray==2.52.1
+referencing==0.37.0
+regex==2025.11.3
+requests==2.32.5
+rich==14.2.0
+rich-toolkit==0.17.0
+rignore==0.7.6
+rpds-py==0.30.0
+rsa==4.9.1
+safetensors==0.7.0
+scikit-learn==1.7.2
+scipy==1.15.3
+sentencepiece==0.2.1
+sentry-sdk==2.46.0
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.17.0
+smart_open==7.5.0
+smmap==5.0.2
+sniffio==1.3.1
+soupsieve==2.8
+spacy==3.8.11
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.2
+sse-starlette==3.0.3
+stack-data==0.6.3
+starlette==0.50.0
+sympy==1.13.1
+tabulate==0.9.0
+tenacity==9.1.2
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+tensordict==0.8.3
+thefuzz==0.22.1
+thinc==8.3.10
+threadpoolctl==3.6.0
+tiktoken==0.12.0
+together==1.5.31
+tokenizers==0.22.1
+tomli==2.3.0
+torch==2.6.0+cu124
+torchaudio==2.6.0
+torchdata==0.11.0
+torchvision==0.21.0
+tqdm==4.67.1
+traitlets==5.14.3
+transformers==4.57.3
+triton==3.2.0
+typer==0.19.2
+typer-slim==0.20.0
+typing_extensions==4.15.0
+typing-inspection==0.4.2
+tzdata==2025.2
+urllib3==2.5.0
+uvicorn==0.38.0
+uvloop==0.22.1
+verl==0.5.0.dev0
+virtualenv==20.35.4
+vllm==0.8.5
+wandb==0.23.0
+wasabi==1.1.3
+watchfiles==1.1.1
+wcwidth==0.2.14
+weasel==0.4.3
+websockets==15.0.1
+Werkzeug==3.1.4
+wheel==0.45.1
+wrapt==2.0.1
+xformers==0.0.29.post2
+xgrammar==0.1.18
+xxhash==3.6.0
+yarl==1.22.0
+zipp==3.23.0
+typeguard==4.4.4
+tyro==1.0.3
+verl==0.5.0.dev0
+autocommand==2.2.2
+backports.tarfile==1.2.0
+importlib_metadata==8.0.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+jaraco.context==5.3.0
+jaraco.functools==4.0.1
+jaraco.text==3.12.1
+more-itertools==10.3.0
+packaging==24.2
+platformdirs==4.2.2
+tomli==2.0.1
+typeguard==4.3.0
+typing_extensions==4.12.2
+wheel==0.45.1
+zipp==3.19.2

wandb/run-20260112_223643-sgt786ss/files/wandb-metadata.json

@@ -0,0 +1,64 @@
+{
+  "os": "Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.19",
+  "startedAt": "2026-01-12T14:36:43.218427Z",
+  "args": [
+    "--track"
+  ],
+  "program": "/mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py",
+  "codePath": "cleanrl/cleanrl/scout_dqn/dqn_sudoku.py",
+  "codePathLocal": "cleanrl/cleanrl/scout_dqn/dqn_sudoku.py",
+  "git": {
+    "remote": "https://github.com/mll-lab-nu/RAGEN.git",
+    "commit": "8d73639b99b38265453f898b8d6af7d4af50d56e"
+  },
+  "email": "haoyu-wa22@mails.tsinghua.edu.cn",
+  "root": "/mnt/general/wanghy/RAGEN",
+  "host": "pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0",
+  "executable": "/root/local/miniconda3/envs/ragen/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "2576980377600",
+      "used": "3724296192"
+    }
+  },
+  "memory": {
+    "total": "2163639439360"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-cff454b9-f118-535b-5e1f-6d503e78e52e"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9e77c178-188c-f869-780e-858c20a7d322"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "shc54txtsp3x3q5ajbrzfnd1sznlb05g"
+}

wandb/run-20260112_223643-sgt786ss/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_step":207700,"global_step":207700,"eval/std_return":6.124492096849747,"train/loss":94802.0234375,"eval/success_rate":0.75925,"_wandb":{"runtime":942},"charts/epsilon":0.506714875,"eval/avg_return":12.523325000000002,"rollout/success":1,"eval/episodes":4000,"_runtime":942.891776839,"train/learning_rate":0.0003,"_timestamp":1.7682295476527374e+09,"rollout/success_rate_100":0.77,"perf/SPS":221}

wandb/run-20260112_223856-d964bhe6/files/code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py

@@ -0,0 +1,454 @@
+# DQN with Step Penalty & Time Limit Bootstrap for RAGEN Rubik's Cube 2x2
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+import re
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+# 假设 ragen 库在当前目录的上两级，请根据实际情况调整
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.rubikscube.env import RubiksCube2x2Env
+from ragen.env.rubikscube.config import RubiksCube2x2Config
+
+
+class RubiksCubeWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: RubiksCube2x2Env):
+        super().__init__()
+        self._env = env
+        self._colors = ['W', 'O', 'G', 'R', 'B', 'Y']
+        self._color_to_idx = {c: i for i, c in enumerate(self._colors)}
+        self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(24 * len(self._colors),), dtype=np.float32)
+        self.action_space = gym.spaces.Discrete(12)
+        self._face_pat = re.compile(r"\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])\n\s*\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])")
+        
+        # [关键修改] 每步惩罚
+        # 如果20步都没解出来，累计惩罚是 20 * -0.05 = -1.0，抵消掉最后可能获得的 +1.0
+        # 这迫使 agent 寻找更短路径
+        self.step_penalty = -0.02
+
+    def _encode_obs(self, text_obs: str) -> np.ndarray:
+        faces_order = ["Up (U):", "Left (L):", "Front (F):", "Right (R):", "Back (B):", "Down (D):"]
+        lines = text_obs.splitlines()
+        blocks: List[str] = []
+        i = 0
+        while i < len(lines):
+            line = lines[i]
+            for header in faces_order:
+                if line.startswith(header):
+                    content = line[len(header):].strip()
+                    next_line = lines[i + 1] if i + 1 < len(lines) else ""
+                    block = f"{content}\n{next_line}"
+                    blocks.append(block)
+                    break
+            i += 1
+        if len(blocks) != 6:
+            return np.zeros(24 * len(self._colors), dtype=np.float32)
+        stickers: List[int] = []
+        for blk in blocks:
+            m = self._face_pat.search(blk)
+            if not m:
+                return np.zeros(24 * len(self._colors), dtype=np.float32)
+            c0, c1, c2, c3 = m.group(1), m.group(2), m.group(3), m.group(4)
+            stickers.extend([c0, c1, c2, c3])
+        grid = np.zeros((24, len(self._colors)), dtype=np.float32)
+        for idx, ch in enumerate(stickers):
+            cidx = self._color_to_idx.get(ch, None)
+            if cidx is not None:
+                grid[idx, cidx] = 1.0
+        return grid.reshape(-1)
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        mapped = int(action) + 1
+        text_obs, reward, done, info = self._env.step(mapped)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        
+        # [关键修改] 应用惩罚
+        # 此时 reward 包含原本的稀疏奖励 (0 或 1) 加上每步惩罚
+        reward = float(reward) + self.step_penalty
+        
+        return obs, reward, terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "RubiksCube2x2DQN"
+    total_timesteps: int = 400_000
+    learning_rate: float = 2.5e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 200_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 4
+    learning_starts: int = 5000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Rubik specific
+    scramble_depth: int = 3 # 初始难度
+    max_steps_env: int = 20 # 硬性时间限制
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, scramble_depth, max_steps_env, capture_video=False):
+    def thunk():
+        # [关键修改] 内部 max_steps 设为 1000，防止 env 内部发出 done=True
+        # 我们完全依赖外部 TimeLimit wrapper 来处理超时
+        config = RubiksCube2x2Config(scramble_depth=scramble_depth, max_steps=20, render_mode='text')
+        env = RubiksCube2x2Env(config)
+        env = RubiksCubeWrapper(env)
+        # 外部限制设为 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps_env)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_shape: Tuple[int, ...]):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.scramble_depth, args.max_steps_env, args.capture_video)()
+    obs_shape = env.observation_space.shape
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_shape)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    # 评估函数
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                state, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                traj_states = [np.asarray(state).tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.max_steps_env)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        action = int(torch.argmax(q, dim=1).item())
+                    next_state, reward, terminated, truncated, info = env_eval.step(action)
+                    
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_state
+                    traj_states.append(np.asarray(state).tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    # 主训练循环
+    global_step = 0
+    start_time = time.time()
+
+    obs, _ = env.reset(seed=args.seed)
+    ep_return = 0.0
+    ep_len = 0
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        if np.random.rand() < epsilon or global_step < args.learning_starts:
+            action = env.action_space.sample()
+        else:
+            with torch.no_grad():
+                q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+                action = int(torch.argmax(q_values, dim=1).item())
+
+        next_obs, reward, terminated, truncated, info = env.step(action)
+        
+        # 判断是否需要 Reset (任何结束都 Reset)
+        real_done = bool(terminated) or bool(truncated)
+
+        # [核心逻辑]
+        # 1. 即使是 truncated (超时)，done 也记为 False，以便进行 bootstrap (计算未来价值)
+        # 2. 只有 terminated (真正解开了)，done 才记为 True
+        # 3. Step Penalty 已经包含在 reward 中，DQN 会学到"为了避免扣分，必须在未来几步内解决"
+        rb.add(
+            obs.astype(np.float32), 
+            int(action), 
+            float(reward), 
+            bool(terminated), # 注意：只用 terminated
+            next_obs.astype(np.float32)
+        )
+
+        obs = next_obs
+        ep_return += float(reward)
+        ep_len += 1
+        global_step += 1
+
+        # 训练过程
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        # 结束时 Reset
+        if real_done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/episodic_return": float(ep_return),
+                        "train/episodic_length": int(ep_len),
+                        "train/success": float(1.0 if succ else 0.0),
+                        "train/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) >= 1 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs, _ = env.reset()
+            ep_return, ep_len = 0.0, 0
+
+        # 更新目标网络
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        # 打印日志
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        # 评估
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    # 评估环境也同样设置：内部1000，外部20
+                    return make_env(0, run_name, args.seed + 9999, args.scramble_depth, args.max_steps_env, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_223856-d964bhe6/files/config.yaml

@@ -0,0 +1,167 @@
+_wandb:
+    value:
+        cli_version: 0.23.0
+        code_path: code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+        e:
+            fv4797ph5nqwn56g5wjpnnfak73gdsee:
+                args:
+                    - --track
+                codePath: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                codePathLocal: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "2576980377600"
+                        used: "3724300288"
+                email: haoyu-wa22@mails.tsinghua.edu.cn
+                executable: /root/local/miniconda3/envs/ragen/bin/python
+                git:
+                    commit: 8d73639b99b38265453f898b8d6af7d4af50d56e
+                    remote: https://github.com/mll-lab-nu/RAGEN.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-cff454b9-f118-535b-5e1f-6d503e78e52e
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9e77c178-188c-f869-780e-858c20a7d322
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7
+                host: pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0
+                memory:
+                    total: "2163639439360"
+                os: Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35
+                program: /mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                python: CPython 3.10.19
+                root: /mnt/general/wanghy/RAGEN
+                startedAt: "2026-01-12T14:38:56.780519Z"
+                writerId: fv4797ph5nqwn56g5wjpnnfak73gdsee
+        m:
+            - "1": global_step
+              "6":
+                - 3
+              "7": []
+            - "2": losses/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": charts/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": perf/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": train/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": rollout/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": eval/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+        python_version: 3.10.19
+        t:
+            "1":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "2":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "3":
+                - 7
+                - 13
+                - 16
+                - 61
+            "4": 3.10.19
+            "5": 0.23.0
+            "6": 4.57.3
+            "12": 0.23.0
+            "13": linux-x86_64
+batch_size:
+    value: 128
+buffer_size:
+    value: 200000
+capture_video:
+    value: false
+cuda:
+    value: true
+end_e:
+    value: 0.05
+env_id:
+    value: RubiksCube2x2DQN
+eval_episodes:
+    value: 4000
+eval_splits:
+    value: 2
+exp_name:
+    value: dqn_rubikscube
+exploration_fraction:
+    value: 0.2
+gamma:
+    value: 0.99
+hidden_size:
+    value: 256
+learning_rate:
+    value: 0.00025
+learning_starts:
+    value: 5000
+max_steps_env:
+    value: 20
+scramble_depth:
+    value: 3
+seed:
+    value: 1
+start_e:
+    value: 1
+target_network_frequency:
+    value: 4000
+torch_deterministic:
+    value: true
+total_timesteps:
+    value: 400000
+track:
+    value: true
+train_frequency:
+    value: 4
+wandb_entity:
+    value: null
+wandb_project_name:
+    value: cleanRL

wandb/run-20260112_223856-d964bhe6/files/diff.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_223856-d964bhe6/files/diff_8d73639b99b38265453f898b8d6af7d4af50d56e.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_223856-d964bhe6/files/requirements.txt

@@ -0,0 +1,313 @@
+colorama==0.4.6
+psutil==7.1.3
+ragen==0.1
+Farama-Notifications==0.0.4
+MarkupSafe==2.1.5
+PyJWT==2.10.1
+PySocks==1.7.1
+absl-py==2.3.1
+accelerate==1.12.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.2
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+airportsdata==20250909
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anthropic==0.75.0
+antlr4-python3-runtime==4.9.3
+anyio==4.12.0
+astor==0.8.1
+asttokens==3.0.1
+async-timeout==5.0.1
+attrs==25.4.0
+beautifulsoup4==4.14.3
+black==25.11.0
+blake3==1.0.8
+blinker==1.9.0
+blis==1.3.3
+cachetools==6.2.2
+catalogue==2.0.10
+certifi==2025.11.12
+cffi==2.0.0
+charset-normalizer==3.4.4
+cleantext==1.1.4
+click==8.2.1
+cloudpathlib==0.23.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+coloredlogs==15.0.1
+colorful==0.5.8
+compressed-tensors==0.9.3
+confection==0.1.5
+contourpy==1.3.2
+cryptography==46.0.3
+cupy-cuda12x==13.6.0
+cycler==0.12.1
+cymem==2.0.13
+Cython==3.2.2
+datasets==4.4.1
+debugpy==1.8.17
+decorator==5.2.1
+Deprecated==1.3.1
+depyf==0.18.0
+dill==0.4.0
+diskcache==5.6.3
+distlib==0.4.0
+distro==1.9.0
+dnspython==2.8.0
+docstring_parser==0.17.0
+einops==0.8.1
+email-validator==2.3.0
+eval_type_backport==0.2.2
+exceptiongroup==1.3.1
+executing==2.2.1
+faiss-gpu==1.7.2
+fastapi==0.123.1
+fastapi-cli==0.0.16
+fastapi-cloud-cli==0.5.2
+fastar==0.8.0
+fastrlock==0.8.3
+filelock==3.19.1
+flash_attn==2.8.3
+Flask==3.1.2
+flatbuffers==25.9.23
+fonttools==4.61.0
+frozenlist==1.8.0
+fsspec==2025.9.0
+gdown==5.2.0
+gguf==0.17.1
+gitdb==4.0.12
+GitPython==3.1.45
+google-api-core==2.28.1
+google-auth==2.43.0
+googleapis-common-protos==1.72.0
+grpcio==1.76.0
+gym==0.26.2
+gym-notices==0.1.0
+gym-sokoban==0.0.6
+gymnasium==1.2.2
+h11==0.16.0
+hf-xet==1.2.0
+html2text==2025.4.15
+httpcore==1.0.9
+httptools==0.7.1
+httpx==0.28.1
+httpx-sse==0.4.3
+huggingface-hub==0.36.0
+humanfriendly==10.0
+hydra-core==1.3.2
+idna==3.11
+ImageIO==2.37.2
+importlib_metadata==8.0.0
+interegular==0.3.3
+ipython==8.37.0
+itsdangerous==2.2.0
+jedi==0.19.2
+Jinja2==3.1.6
+jiter==0.12.0
+joblib==1.5.2
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+kimina-client==0.2.1
+kiwisolver==1.4.9
+lark==1.2.2
+llguidance==0.7.30
+llvmlite==0.44.0
+lm-format-enforcer==0.10.12
+loguru==0.7.3
+Markdown==3.10
+markdown-it-py==4.0.0
+matplotlib==3.10.7
+matplotlib-inline==0.2.1
+mcp==1.22.0
+mdurl==0.1.2
+mistral_common==1.8.6
+mpmath==1.3.0
+msgpack==1.1.2
+msgspec==0.20.0
+multidict==6.7.0
+multiprocess==0.70.18
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+nest-asyncio==1.6.0
+networkx==3.3
+ninja==1.13.0
+nltk==3.9.2
+numba==0.61.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-cusparselt-cu12==0.6.2
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+omegaconf==2.3.0
+onnxruntime==1.23.2
+openai==2.8.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opencv-python-headless==4.12.0.88
+opentelemetry-api==1.38.0
+opentelemetry-exporter-otlp==1.26.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+opentelemetry-exporter-prometheus==0.59b0
+opentelemetry-proto==1.26.0
+opentelemetry-sdk==1.38.0
+opentelemetry-semantic-conventions==0.59b0
+opentelemetry-semantic-conventions-ai==0.4.13
+orjson==3.11.4
+outlines==0.1.11
+outlines_core==0.1.26
+packaging==25.0
+pandas==2.3.3
+parso==0.8.5
+partial-json-parser==0.2.1.1.post7
+pathspec==0.12.1
+peft==0.18.0
+pexpect==4.9.0
+pillow==11.3.0
+pip==25.3
+platformdirs==4.5.0
+preshed==3.0.12
+prometheus_client==0.23.1
+prometheus-fastapi-instrumentator==7.1.0
+prompt_toolkit==3.0.52
+propcache==0.4.1
+proto-plus==1.26.1
+protobuf==4.25.8
+psutil==7.1.3
+ptyprocess==0.7.0
+pure_eval==0.2.3
+py-cpuinfo==9.0.0
+py-spy==0.4.1
+pyarrow==22.0.0
+pyasn1==0.6.1
+pyasn1_modules==0.4.2
+pybind11==3.0.1
+pycountry==24.6.1
+pycparser==2.23
+pydantic==2.12.5
+pydantic_core==2.41.5
+pydantic-extra-types==2.10.6
+pydantic-settings==2.12.0
+pygame==2.6.1
+Pygments==2.19.2
+pyjnius==1.7.0
+pylatexenc==2.10
+pyparsing==3.2.5
+pyserini==1.2.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.1
+python-json-logger==4.0.0
+python-multipart==0.0.20
+pytokens==0.3.0
+pytz==2025.2
+PyYAML==6.0.3
+pyzmq==27.1.0
+ragen==0.1
+rank-bm25==0.2.2
+RapidFuzz==3.14.3
+ray==2.52.1
+referencing==0.37.0
+regex==2025.11.3
+requests==2.32.5
+rich==14.2.0
+rich-toolkit==0.17.0
+rignore==0.7.6
+rpds-py==0.30.0
+rsa==4.9.1
+safetensors==0.7.0
+scikit-learn==1.7.2
+scipy==1.15.3
+sentencepiece==0.2.1
+sentry-sdk==2.46.0
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.17.0
+smart_open==7.5.0
+smmap==5.0.2
+sniffio==1.3.1
+soupsieve==2.8
+spacy==3.8.11
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.2
+sse-starlette==3.0.3
+stack-data==0.6.3
+starlette==0.50.0
+sympy==1.13.1
+tabulate==0.9.0
+tenacity==9.1.2
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+tensordict==0.8.3
+thefuzz==0.22.1
+thinc==8.3.10
+threadpoolctl==3.6.0
+tiktoken==0.12.0
+together==1.5.31
+tokenizers==0.22.1
+tomli==2.3.0
+torch==2.6.0+cu124
+torchaudio==2.6.0
+torchdata==0.11.0
+torchvision==0.21.0
+tqdm==4.67.1
+traitlets==5.14.3
+transformers==4.57.3
+triton==3.2.0
+typer==0.19.2
+typer-slim==0.20.0
+typing_extensions==4.15.0
+typing-inspection==0.4.2
+tzdata==2025.2
+urllib3==2.5.0
+uvicorn==0.38.0
+uvloop==0.22.1
+verl==0.5.0.dev0
+virtualenv==20.35.4
+vllm==0.8.5
+wandb==0.23.0
+wasabi==1.1.3
+watchfiles==1.1.1
+wcwidth==0.2.14
+weasel==0.4.3
+websockets==15.0.1
+Werkzeug==3.1.4
+wheel==0.45.1
+wrapt==2.0.1
+xformers==0.0.29.post2
+xgrammar==0.1.18
+xxhash==3.6.0
+yarl==1.22.0
+zipp==3.23.0
+typeguard==4.4.4
+tyro==1.0.3
+verl==0.5.0.dev0
+autocommand==2.2.2
+backports.tarfile==1.2.0
+importlib_metadata==8.0.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+jaraco.context==5.3.0
+jaraco.functools==4.0.1
+jaraco.text==3.12.1
+more-itertools==10.3.0
+packaging==24.2
+platformdirs==4.2.2
+tomli==2.0.1
+typeguard==4.3.0
+typing_extensions==4.12.2
+wheel==0.45.1
+zipp==3.19.2

wandb/run-20260112_223856-d964bhe6/files/wandb-metadata.json

@@ -0,0 +1,64 @@
+{
+  "os": "Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.19",
+  "startedAt": "2026-01-12T14:38:56.780519Z",
+  "args": [
+    "--track"
+  ],
+  "program": "/mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePath": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePathLocal": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "git": {
+    "remote": "https://github.com/mll-lab-nu/RAGEN.git",
+    "commit": "8d73639b99b38265453f898b8d6af7d4af50d56e"
+  },
+  "email": "haoyu-wa22@mails.tsinghua.edu.cn",
+  "root": "/mnt/general/wanghy/RAGEN",
+  "host": "pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0",
+  "executable": "/root/local/miniconda3/envs/ragen/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "2576980377600",
+      "used": "3724300288"
+    }
+  },
+  "memory": {
+    "total": "2163639439360"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-cff454b9-f118-535b-5e1f-6d503e78e52e"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9e77c178-188c-f869-780e-858c20a7d322"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "fv4797ph5nqwn56g5wjpnnfak73gdsee"
+}

wandb/run-20260112_223856-d964bhe6/files/wandb-summary.json

@@ -0,0 +1 @@
+{"charts/epsilon":0.05,"train/episodic_return":-0.4000000000000001,"_timestamp":1.7682289096790524e+09,"train/success_rate_100":0.04,"train/loss":0.012717867270112038,"train/success":0,"global_step":99692,"_step":99692,"eval/avg_return":-0.3364750000000001,"eval/success_rate":0.04625,"train/learning_rate":0.00025,"_runtime":171.599183111,"train/episodic_length":20,"perf/SPS":596,"_wandb":{"runtime":171},"eval/episodes":4000,"eval/std_return":0.2884910299732038}

wandb/run-20260112_224209-62o99t0u/files/code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py

@@ -0,0 +1,454 @@
+# DQN with Step Penalty & Time Limit Bootstrap for RAGEN Rubik's Cube 2x2
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+import re
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+# 假设 ragen 库在当前目录的上两级，请根据实际情况调整
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.rubikscube.env import RubiksCube2x2Env
+from ragen.env.rubikscube.config import RubiksCube2x2Config
+
+
+class RubiksCubeWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: RubiksCube2x2Env):
+        super().__init__()
+        self._env = env
+        self._colors = ['W', 'O', 'G', 'R', 'B', 'Y']
+        self._color_to_idx = {c: i for i, c in enumerate(self._colors)}
+        self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(24 * len(self._colors),), dtype=np.float32)
+        self.action_space = gym.spaces.Discrete(12)
+        self._face_pat = re.compile(r"\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])\n\s*\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])")
+        
+        # [关键修改] 每步惩罚
+        # 如果20步都没解出来，累计惩罚是 20 * -0.05 = -1.0，抵消掉最后可能获得的 +1.0
+        # 这迫使 agent 寻找更短路径
+        self.step_penalty = -0.045
+
+    def _encode_obs(self, text_obs: str) -> np.ndarray:
+        faces_order = ["Up (U):", "Left (L):", "Front (F):", "Right (R):", "Back (B):", "Down (D):"]
+        lines = text_obs.splitlines()
+        blocks: List[str] = []
+        i = 0
+        while i < len(lines):
+            line = lines[i]
+            for header in faces_order:
+                if line.startswith(header):
+                    content = line[len(header):].strip()
+                    next_line = lines[i + 1] if i + 1 < len(lines) else ""
+                    block = f"{content}\n{next_line}"
+                    blocks.append(block)
+                    break
+            i += 1
+        if len(blocks) != 6:
+            return np.zeros(24 * len(self._colors), dtype=np.float32)
+        stickers: List[int] = []
+        for blk in blocks:
+            m = self._face_pat.search(blk)
+            if not m:
+                return np.zeros(24 * len(self._colors), dtype=np.float32)
+            c0, c1, c2, c3 = m.group(1), m.group(2), m.group(3), m.group(4)
+            stickers.extend([c0, c1, c2, c3])
+        grid = np.zeros((24, len(self._colors)), dtype=np.float32)
+        for idx, ch in enumerate(stickers):
+            cidx = self._color_to_idx.get(ch, None)
+            if cidx is not None:
+                grid[idx, cidx] = 1.0
+        return grid.reshape(-1)
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        mapped = int(action) + 1
+        text_obs, reward, done, info = self._env.step(mapped)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        
+        # [关键修改] 应用惩罚
+        # 此时 reward 包含原本的稀疏奖励 (0 或 1) 加上每步惩罚
+        reward = float(reward) + self.step_penalty
+        
+        return obs, reward, terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "RubiksCube2x2DQN"
+    total_timesteps: int = 400_000
+    learning_rate: float = 2.5e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 200_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 4
+    learning_starts: int = 5000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Rubik specific
+    scramble_depth: int = 3 # 初始难度
+    max_steps_env: int = 20 # 硬性时间限制
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, scramble_depth, max_steps_env, capture_video=False):
+    def thunk():
+        # [关键修改] 内部 max_steps 设为 1000，防止 env 内部发出 done=True
+        # 我们完全依赖外部 TimeLimit wrapper 来处理超时
+        config = RubiksCube2x2Config(scramble_depth=scramble_depth, max_steps=20, render_mode='text')
+        env = RubiksCube2x2Env(config)
+        env = RubiksCubeWrapper(env)
+        # 外部限制设为 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps_env)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_shape: Tuple[int, ...]):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.scramble_depth, args.max_steps_env, args.capture_video)()
+    obs_shape = env.observation_space.shape
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_shape)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    # 评估函数
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                state, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                traj_states = [np.asarray(state).tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.max_steps_env)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        action = int(torch.argmax(q, dim=1).item())
+                    next_state, reward, terminated, truncated, info = env_eval.step(action)
+                    
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_state
+                    traj_states.append(np.asarray(state).tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    # 主训练循环
+    global_step = 0
+    start_time = time.time()
+
+    obs, _ = env.reset(seed=args.seed)
+    ep_return = 0.0
+    ep_len = 0
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        if np.random.rand() < epsilon or global_step < args.learning_starts:
+            action = env.action_space.sample()
+        else:
+            with torch.no_grad():
+                q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+                action = int(torch.argmax(q_values, dim=1).item())
+
+        next_obs, reward, terminated, truncated, info = env.step(action)
+        
+        # 判断是否需要 Reset (任何结束都 Reset)
+        real_done = bool(terminated) or bool(truncated)
+
+        # [核心逻辑]
+        # 1. 即使是 truncated (超时)，done 也记为 False，以便进行 bootstrap (计算未来价值)
+        # 2. 只有 terminated (真正解开了)，done 才记为 True
+        # 3. Step Penalty 已经包含在 reward 中，DQN 会学到"为了避免扣分，必须在未来几步内解决"
+        rb.add(
+            obs.astype(np.float32), 
+            int(action), 
+            float(reward), 
+            bool(terminated), # 注意：只用 terminated
+            next_obs.astype(np.float32)
+        )
+
+        obs = next_obs
+        ep_return += float(reward)
+        ep_len += 1
+        global_step += 1
+
+        # 训练过程
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        # 结束时 Reset
+        if real_done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/episodic_return": float(ep_return),
+                        "train/episodic_length": int(ep_len),
+                        "train/success": float(1.0 if succ else 0.0),
+                        "train/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) >= 1 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs, _ = env.reset()
+            ep_return, ep_len = 0.0, 0
+
+        # 更新目标网络
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        # 打印日志
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        # 评估
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    # 评估环境也同样设置：内部1000，外部20
+                    return make_env(0, run_name, args.seed + 9999, args.scramble_depth, args.max_steps_env, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_224209-62o99t0u/files/config.yaml

@@ -0,0 +1,167 @@
+_wandb:
+    value:
+        cli_version: 0.23.0
+        code_path: code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+        e:
+            35ar29uupah1ktpgjn4djl5xx5ow78o3:
+                args:
+                    - --track
+                codePath: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                codePathLocal: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "2576980377600"
+                        used: "3724304384"
+                email: haoyu-wa22@mails.tsinghua.edu.cn
+                executable: /root/local/miniconda3/envs/ragen/bin/python
+                git:
+                    commit: 8d73639b99b38265453f898b8d6af7d4af50d56e
+                    remote: https://github.com/mll-lab-nu/RAGEN.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-cff454b9-f118-535b-5e1f-6d503e78e52e
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9e77c178-188c-f869-780e-858c20a7d322
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7
+                host: pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0
+                memory:
+                    total: "2163639439360"
+                os: Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35
+                program: /mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                python: CPython 3.10.19
+                root: /mnt/general/wanghy/RAGEN
+                startedAt: "2026-01-12T14:42:09.125131Z"
+                writerId: 35ar29uupah1ktpgjn4djl5xx5ow78o3
+        m:
+            - "1": global_step
+              "6":
+                - 3
+              "7": []
+            - "2": perf/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": train/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": rollout/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": eval/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": losses/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": charts/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+        python_version: 3.10.19
+        t:
+            "1":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "2":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "3":
+                - 7
+                - 13
+                - 16
+                - 61
+            "4": 3.10.19
+            "5": 0.23.0
+            "6": 4.57.3
+            "12": 0.23.0
+            "13": linux-x86_64
+batch_size:
+    value: 128
+buffer_size:
+    value: 200000
+capture_video:
+    value: false
+cuda:
+    value: true
+end_e:
+    value: 0.05
+env_id:
+    value: RubiksCube2x2DQN
+eval_episodes:
+    value: 4000
+eval_splits:
+    value: 2
+exp_name:
+    value: dqn_rubikscube
+exploration_fraction:
+    value: 0.2
+gamma:
+    value: 0.99
+hidden_size:
+    value: 256
+learning_rate:
+    value: 0.00025
+learning_starts:
+    value: 5000
+max_steps_env:
+    value: 20
+scramble_depth:
+    value: 3
+seed:
+    value: 1
+start_e:
+    value: 1
+target_network_frequency:
+    value: 4000
+torch_deterministic:
+    value: true
+total_timesteps:
+    value: 400000
+track:
+    value: true
+train_frequency:
+    value: 4
+wandb_entity:
+    value: null
+wandb_project_name:
+    value: cleanRL

wandb/run-20260112_224209-62o99t0u/files/diff.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_224209-62o99t0u/files/diff_8d73639b99b38265453f898b8d6af7d4af50d56e.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_224209-62o99t0u/files/requirements.txt

@@ -0,0 +1,313 @@
+colorama==0.4.6
+psutil==7.1.3
+ragen==0.1
+Farama-Notifications==0.0.4
+MarkupSafe==2.1.5
+PyJWT==2.10.1
+PySocks==1.7.1
+absl-py==2.3.1
+accelerate==1.12.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.2
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+airportsdata==20250909
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anthropic==0.75.0
+antlr4-python3-runtime==4.9.3
+anyio==4.12.0
+astor==0.8.1
+asttokens==3.0.1
+async-timeout==5.0.1
+attrs==25.4.0
+beautifulsoup4==4.14.3
+black==25.11.0
+blake3==1.0.8
+blinker==1.9.0
+blis==1.3.3
+cachetools==6.2.2
+catalogue==2.0.10
+certifi==2025.11.12
+cffi==2.0.0
+charset-normalizer==3.4.4
+cleantext==1.1.4
+click==8.2.1
+cloudpathlib==0.23.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+coloredlogs==15.0.1
+colorful==0.5.8
+compressed-tensors==0.9.3
+confection==0.1.5
+contourpy==1.3.2
+cryptography==46.0.3
+cupy-cuda12x==13.6.0
+cycler==0.12.1
+cymem==2.0.13
+Cython==3.2.2
+datasets==4.4.1
+debugpy==1.8.17
+decorator==5.2.1
+Deprecated==1.3.1
+depyf==0.18.0
+dill==0.4.0
+diskcache==5.6.3
+distlib==0.4.0
+distro==1.9.0
+dnspython==2.8.0
+docstring_parser==0.17.0
+einops==0.8.1
+email-validator==2.3.0
+eval_type_backport==0.2.2
+exceptiongroup==1.3.1
+executing==2.2.1
+faiss-gpu==1.7.2
+fastapi==0.123.1
+fastapi-cli==0.0.16
+fastapi-cloud-cli==0.5.2
+fastar==0.8.0
+fastrlock==0.8.3
+filelock==3.19.1
+flash_attn==2.8.3
+Flask==3.1.2
+flatbuffers==25.9.23
+fonttools==4.61.0
+frozenlist==1.8.0
+fsspec==2025.9.0
+gdown==5.2.0
+gguf==0.17.1
+gitdb==4.0.12
+GitPython==3.1.45
+google-api-core==2.28.1
+google-auth==2.43.0
+googleapis-common-protos==1.72.0
+grpcio==1.76.0
+gym==0.26.2
+gym-notices==0.1.0
+gym-sokoban==0.0.6
+gymnasium==1.2.2
+h11==0.16.0
+hf-xet==1.2.0
+html2text==2025.4.15
+httpcore==1.0.9
+httptools==0.7.1
+httpx==0.28.1
+httpx-sse==0.4.3
+huggingface-hub==0.36.0
+humanfriendly==10.0
+hydra-core==1.3.2
+idna==3.11
+ImageIO==2.37.2
+importlib_metadata==8.0.0
+interegular==0.3.3
+ipython==8.37.0
+itsdangerous==2.2.0
+jedi==0.19.2
+Jinja2==3.1.6
+jiter==0.12.0
+joblib==1.5.2
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+kimina-client==0.2.1
+kiwisolver==1.4.9
+lark==1.2.2
+llguidance==0.7.30
+llvmlite==0.44.0
+lm-format-enforcer==0.10.12
+loguru==0.7.3
+Markdown==3.10
+markdown-it-py==4.0.0
+matplotlib==3.10.7
+matplotlib-inline==0.2.1
+mcp==1.22.0
+mdurl==0.1.2
+mistral_common==1.8.6
+mpmath==1.3.0
+msgpack==1.1.2
+msgspec==0.20.0
+multidict==6.7.0
+multiprocess==0.70.18
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+nest-asyncio==1.6.0
+networkx==3.3
+ninja==1.13.0
+nltk==3.9.2
+numba==0.61.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-cusparselt-cu12==0.6.2
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+omegaconf==2.3.0
+onnxruntime==1.23.2
+openai==2.8.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opencv-python-headless==4.12.0.88
+opentelemetry-api==1.38.0
+opentelemetry-exporter-otlp==1.26.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+opentelemetry-exporter-prometheus==0.59b0
+opentelemetry-proto==1.26.0
+opentelemetry-sdk==1.38.0
+opentelemetry-semantic-conventions==0.59b0
+opentelemetry-semantic-conventions-ai==0.4.13
+orjson==3.11.4
+outlines==0.1.11
+outlines_core==0.1.26
+packaging==25.0
+pandas==2.3.3
+parso==0.8.5
+partial-json-parser==0.2.1.1.post7
+pathspec==0.12.1
+peft==0.18.0
+pexpect==4.9.0
+pillow==11.3.0
+pip==25.3
+platformdirs==4.5.0
+preshed==3.0.12
+prometheus_client==0.23.1
+prometheus-fastapi-instrumentator==7.1.0
+prompt_toolkit==3.0.52
+propcache==0.4.1
+proto-plus==1.26.1
+protobuf==4.25.8
+psutil==7.1.3
+ptyprocess==0.7.0
+pure_eval==0.2.3
+py-cpuinfo==9.0.0
+py-spy==0.4.1
+pyarrow==22.0.0
+pyasn1==0.6.1
+pyasn1_modules==0.4.2
+pybind11==3.0.1
+pycountry==24.6.1
+pycparser==2.23
+pydantic==2.12.5
+pydantic_core==2.41.5
+pydantic-extra-types==2.10.6
+pydantic-settings==2.12.0
+pygame==2.6.1
+Pygments==2.19.2
+pyjnius==1.7.0
+pylatexenc==2.10
+pyparsing==3.2.5
+pyserini==1.2.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.1
+python-json-logger==4.0.0
+python-multipart==0.0.20
+pytokens==0.3.0
+pytz==2025.2
+PyYAML==6.0.3
+pyzmq==27.1.0
+ragen==0.1
+rank-bm25==0.2.2
+RapidFuzz==3.14.3
+ray==2.52.1
+referencing==0.37.0
+regex==2025.11.3
+requests==2.32.5
+rich==14.2.0
+rich-toolkit==0.17.0
+rignore==0.7.6
+rpds-py==0.30.0
+rsa==4.9.1
+safetensors==0.7.0
+scikit-learn==1.7.2
+scipy==1.15.3
+sentencepiece==0.2.1
+sentry-sdk==2.46.0
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.17.0
+smart_open==7.5.0
+smmap==5.0.2
+sniffio==1.3.1
+soupsieve==2.8
+spacy==3.8.11
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.2
+sse-starlette==3.0.3
+stack-data==0.6.3
+starlette==0.50.0
+sympy==1.13.1
+tabulate==0.9.0
+tenacity==9.1.2
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+tensordict==0.8.3
+thefuzz==0.22.1
+thinc==8.3.10
+threadpoolctl==3.6.0
+tiktoken==0.12.0
+together==1.5.31
+tokenizers==0.22.1
+tomli==2.3.0
+torch==2.6.0+cu124
+torchaudio==2.6.0
+torchdata==0.11.0
+torchvision==0.21.0
+tqdm==4.67.1
+traitlets==5.14.3
+transformers==4.57.3
+triton==3.2.0
+typer==0.19.2
+typer-slim==0.20.0
+typing_extensions==4.15.0
+typing-inspection==0.4.2
+tzdata==2025.2
+urllib3==2.5.0
+uvicorn==0.38.0
+uvloop==0.22.1
+verl==0.5.0.dev0
+virtualenv==20.35.4
+vllm==0.8.5
+wandb==0.23.0
+wasabi==1.1.3
+watchfiles==1.1.1
+wcwidth==0.2.14
+weasel==0.4.3
+websockets==15.0.1
+Werkzeug==3.1.4
+wheel==0.45.1
+wrapt==2.0.1
+xformers==0.0.29.post2
+xgrammar==0.1.18
+xxhash==3.6.0
+yarl==1.22.0
+zipp==3.23.0
+typeguard==4.4.4
+tyro==1.0.3
+verl==0.5.0.dev0
+autocommand==2.2.2
+backports.tarfile==1.2.0
+importlib_metadata==8.0.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+jaraco.context==5.3.0
+jaraco.functools==4.0.1
+jaraco.text==3.12.1
+more-itertools==10.3.0
+packaging==24.2
+platformdirs==4.2.2
+tomli==2.0.1
+typeguard==4.3.0
+typing_extensions==4.12.2
+wheel==0.45.1
+zipp==3.19.2

wandb/run-20260112_224209-62o99t0u/files/wandb-metadata.json

@@ -0,0 +1,64 @@
+{
+  "os": "Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.19",
+  "startedAt": "2026-01-12T14:42:09.125131Z",
+  "args": [
+    "--track"
+  ],
+  "program": "/mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePath": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePathLocal": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "git": {
+    "remote": "https://github.com/mll-lab-nu/RAGEN.git",
+    "commit": "8d73639b99b38265453f898b8d6af7d4af50d56e"
+  },
+  "email": "haoyu-wa22@mails.tsinghua.edu.cn",
+  "root": "/mnt/general/wanghy/RAGEN",
+  "host": "pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0",
+  "executable": "/root/local/miniconda3/envs/ragen/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "2576980377600",
+      "used": "3724304384"
+    }
+  },
+  "memory": {
+    "total": "2163639439360"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-cff454b9-f118-535b-5e1f-6d503e78e52e"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9e77c178-188c-f869-780e-858c20a7d322"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "35ar29uupah1ktpgjn4djl5xx5ow78o3"
+}

wandb/run-20260112_224209-62o99t0u/files/wandb-summary.json

@@ -0,0 +1 @@
+{"eval/std_return":0.4499041483176584,"_timestamp":1.768229356683353e+09,"_runtime":426.228488766,"_wandb":{"runtime":426},"eval/episodes":4000,"global_step":252356,"train/episodic_length":1,"train/learning_rate":0.00025,"_step":252356,"eval/success_rate":0.93325,"eval/avg_return":0.77420875,"train/success":1,"perf/SPS":598,"train/episodic_return":0.955,"charts/epsilon":0.05,"train/loss":0.0037821633741259575,"train/success_rate_100":0.97}

wandb/run-20260112_224942-jlcvlgih/files/code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py

@@ -0,0 +1,454 @@
+# DQN with Step Penalty & Time Limit Bootstrap for RAGEN Rubik's Cube 2x2
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+import re
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+# 假设 ragen 库在当前目录的上两级，请根据实际情况调整
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.rubikscube.env import RubiksCube2x2Env
+from ragen.env.rubikscube.config import RubiksCube2x2Config
+
+
+class RubiksCubeWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: RubiksCube2x2Env):
+        super().__init__()
+        self._env = env
+        self._colors = ['W', 'O', 'G', 'R', 'B', 'Y']
+        self._color_to_idx = {c: i for i, c in enumerate(self._colors)}
+        self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(24 * len(self._colors),), dtype=np.float32)
+        self.action_space = gym.spaces.Discrete(12)
+        self._face_pat = re.compile(r"\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])\n\s*\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])")
+        
+        # [关键修改] 每步惩罚
+        # 如果20步都没解出来，累计惩罚是 20 * -0.05 = -1.0，抵消掉最后可能获得的 +1.0
+        # 这迫使 agent 寻找更短路径
+        self.step_penalty = -0.04
+
+    def _encode_obs(self, text_obs: str) -> np.ndarray:
+        faces_order = ["Up (U):", "Left (L):", "Front (F):", "Right (R):", "Back (B):", "Down (D):"]
+        lines = text_obs.splitlines()
+        blocks: List[str] = []
+        i = 0
+        while i < len(lines):
+            line = lines[i]
+            for header in faces_order:
+                if line.startswith(header):
+                    content = line[len(header):].strip()
+                    next_line = lines[i + 1] if i + 1 < len(lines) else ""
+                    block = f"{content}\n{next_line}"
+                    blocks.append(block)
+                    break
+            i += 1
+        if len(blocks) != 6:
+            return np.zeros(24 * len(self._colors), dtype=np.float32)
+        stickers: List[int] = []
+        for blk in blocks:
+            m = self._face_pat.search(blk)
+            if not m:
+                return np.zeros(24 * len(self._colors), dtype=np.float32)
+            c0, c1, c2, c3 = m.group(1), m.group(2), m.group(3), m.group(4)
+            stickers.extend([c0, c1, c2, c3])
+        grid = np.zeros((24, len(self._colors)), dtype=np.float32)
+        for idx, ch in enumerate(stickers):
+            cidx = self._color_to_idx.get(ch, None)
+            if cidx is not None:
+                grid[idx, cidx] = 1.0
+        return grid.reshape(-1)
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        mapped = int(action) + 1
+        text_obs, reward, done, info = self._env.step(mapped)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        
+        # [关键修改] 应用惩罚
+        # 此时 reward 包含原本的稀疏奖励 (0 或 1) 加上每步惩罚
+        reward = float(reward) + self.step_penalty
+        
+        return obs, reward, terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "RubiksCube2x2DQN"
+    total_timesteps: int = 400_000
+    learning_rate: float = 2.5e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 200_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 4
+    learning_starts: int = 5000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Rubik specific
+    scramble_depth: int = 3 # 初始难度
+    max_steps_env: int = 20 # 硬性时间限制
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, scramble_depth, max_steps_env, capture_video=False):
+    def thunk():
+        # [关键修改] 内部 max_steps 设为 1000，防止 env 内部发出 done=True
+        # 我们完全依赖外部 TimeLimit wrapper 来处理超时
+        config = RubiksCube2x2Config(scramble_depth=scramble_depth, max_steps=20, render_mode='text')
+        env = RubiksCube2x2Env(config)
+        env = RubiksCubeWrapper(env)
+        # 外部限制设为 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps_env)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_shape: Tuple[int, ...]):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.scramble_depth, args.max_steps_env, args.capture_video)()
+    obs_shape = env.observation_space.shape
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_shape)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    # 评估函数
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                state, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                traj_states = [np.asarray(state).tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.max_steps_env)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        action = int(torch.argmax(q, dim=1).item())
+                    next_state, reward, terminated, truncated, info = env_eval.step(action)
+                    
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_state
+                    traj_states.append(np.asarray(state).tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    # 主训练循环
+    global_step = 0
+    start_time = time.time()
+
+    obs, _ = env.reset(seed=args.seed)
+    ep_return = 0.0
+    ep_len = 0
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        if np.random.rand() < epsilon or global_step < args.learning_starts:
+            action = env.action_space.sample()
+        else:
+            with torch.no_grad():
+                q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+                action = int(torch.argmax(q_values, dim=1).item())
+
+        next_obs, reward, terminated, truncated, info = env.step(action)
+        
+        # 判断是否需要 Reset (任何结束都 Reset)
+        real_done = bool(terminated) or bool(truncated)
+
+        # [核心逻辑]
+        # 1. 即使是 truncated (超时)，done 也记为 False，以便进行 bootstrap (计算未来价值)
+        # 2. 只有 terminated (真正解开了)，done 才记为 True
+        # 3. Step Penalty 已经包含在 reward 中，DQN 会学到"为了避免扣分，必须在未来几步内解决"
+        rb.add(
+            obs.astype(np.float32), 
+            int(action), 
+            float(reward), 
+            bool(terminated), # 注意：只用 terminated
+            next_obs.astype(np.float32)
+        )
+
+        obs = next_obs
+        ep_return += float(reward)
+        ep_len += 1
+        global_step += 1
+
+        # 训练过程
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        # 结束时 Reset
+        if real_done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/episodic_return": float(ep_return),
+                        "train/episodic_length": int(ep_len),
+                        "train/success": float(1.0 if succ else 0.0),
+                        "train/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) >= 1 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs, _ = env.reset()
+            ep_return, ep_len = 0.0, 0
+
+        # 更新目标网络
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        # 打印日志
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        # 评估
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    # 评估环境也同样设置：内部1000，外部20
+                    return make_env(0, run_name, args.seed + 9999, args.scramble_depth, args.max_steps_env, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_224942-jlcvlgih/files/config.yaml

@@ -0,0 +1,167 @@
+_wandb:
+    value:
+        cli_version: 0.23.0
+        code_path: code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+        e:
+            x6hdflogtnk75ni5ehl9wyx5sede6lv4:
+                args:
+                    - --track
+                codePath: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                codePathLocal: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "2576980377600"
+                        used: "3724308480"
+                email: haoyu-wa22@mails.tsinghua.edu.cn
+                executable: /root/local/miniconda3/envs/ragen/bin/python
+                git:
+                    commit: 8d73639b99b38265453f898b8d6af7d4af50d56e
+                    remote: https://github.com/mll-lab-nu/RAGEN.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-cff454b9-f118-535b-5e1f-6d503e78e52e
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9e77c178-188c-f869-780e-858c20a7d322
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7
+                host: pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0
+                memory:
+                    total: "2163639439360"
+                os: Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35
+                program: /mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                python: CPython 3.10.19
+                root: /mnt/general/wanghy/RAGEN
+                startedAt: "2026-01-12T14:49:42.621345Z"
+                writerId: x6hdflogtnk75ni5ehl9wyx5sede6lv4
+        m:
+            - "1": global_step
+              "6":
+                - 3
+              "7": []
+            - "2": perf/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": train/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": rollout/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": eval/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": losses/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": charts/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+        python_version: 3.10.19
+        t:
+            "1":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "2":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "3":
+                - 7
+                - 13
+                - 16
+                - 61
+            "4": 3.10.19
+            "5": 0.23.0
+            "6": 4.57.3
+            "12": 0.23.0
+            "13": linux-x86_64
+batch_size:
+    value: 128
+buffer_size:
+    value: 200000
+capture_video:
+    value: false
+cuda:
+    value: true
+end_e:
+    value: 0.05
+env_id:
+    value: RubiksCube2x2DQN
+eval_episodes:
+    value: 4000
+eval_splits:
+    value: 2
+exp_name:
+    value: dqn_rubikscube
+exploration_fraction:
+    value: 0.2
+gamma:
+    value: 0.99
+hidden_size:
+    value: 256
+learning_rate:
+    value: 0.00025
+learning_starts:
+    value: 5000
+max_steps_env:
+    value: 20
+scramble_depth:
+    value: 3
+seed:
+    value: 1
+start_e:
+    value: 1
+target_network_frequency:
+    value: 4000
+torch_deterministic:
+    value: true
+total_timesteps:
+    value: 400000
+track:
+    value: true
+train_frequency:
+    value: 4
+wandb_entity:
+    value: null
+wandb_project_name:
+    value: cleanRL

wandb/run-20260112_224942-jlcvlgih/files/diff.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_224942-jlcvlgih/files/diff_8d73639b99b38265453f898b8d6af7d4af50d56e.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_224942-jlcvlgih/files/requirements.txt

@@ -0,0 +1,313 @@
+colorama==0.4.6
+psutil==7.1.3
+ragen==0.1
+Farama-Notifications==0.0.4
+MarkupSafe==2.1.5
+PyJWT==2.10.1
+PySocks==1.7.1
+absl-py==2.3.1
+accelerate==1.12.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.2
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+airportsdata==20250909
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anthropic==0.75.0
+antlr4-python3-runtime==4.9.3
+anyio==4.12.0
+astor==0.8.1
+asttokens==3.0.1
+async-timeout==5.0.1
+attrs==25.4.0
+beautifulsoup4==4.14.3
+black==25.11.0
+blake3==1.0.8
+blinker==1.9.0
+blis==1.3.3
+cachetools==6.2.2
+catalogue==2.0.10
+certifi==2025.11.12
+cffi==2.0.0
+charset-normalizer==3.4.4
+cleantext==1.1.4
+click==8.2.1
+cloudpathlib==0.23.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+coloredlogs==15.0.1
+colorful==0.5.8
+compressed-tensors==0.9.3
+confection==0.1.5
+contourpy==1.3.2
+cryptography==46.0.3
+cupy-cuda12x==13.6.0
+cycler==0.12.1
+cymem==2.0.13
+Cython==3.2.2
+datasets==4.4.1
+debugpy==1.8.17
+decorator==5.2.1
+Deprecated==1.3.1
+depyf==0.18.0
+dill==0.4.0
+diskcache==5.6.3
+distlib==0.4.0
+distro==1.9.0
+dnspython==2.8.0
+docstring_parser==0.17.0
+einops==0.8.1
+email-validator==2.3.0
+eval_type_backport==0.2.2
+exceptiongroup==1.3.1
+executing==2.2.1
+faiss-gpu==1.7.2
+fastapi==0.123.1
+fastapi-cli==0.0.16
+fastapi-cloud-cli==0.5.2
+fastar==0.8.0
+fastrlock==0.8.3
+filelock==3.19.1
+flash_attn==2.8.3
+Flask==3.1.2
+flatbuffers==25.9.23
+fonttools==4.61.0
+frozenlist==1.8.0
+fsspec==2025.9.0
+gdown==5.2.0
+gguf==0.17.1
+gitdb==4.0.12
+GitPython==3.1.45
+google-api-core==2.28.1
+google-auth==2.43.0
+googleapis-common-protos==1.72.0
+grpcio==1.76.0
+gym==0.26.2
+gym-notices==0.1.0
+gym-sokoban==0.0.6
+gymnasium==1.2.2
+h11==0.16.0
+hf-xet==1.2.0
+html2text==2025.4.15
+httpcore==1.0.9
+httptools==0.7.1
+httpx==0.28.1
+httpx-sse==0.4.3
+huggingface-hub==0.36.0
+humanfriendly==10.0
+hydra-core==1.3.2
+idna==3.11
+ImageIO==2.37.2
+importlib_metadata==8.0.0
+interegular==0.3.3
+ipython==8.37.0
+itsdangerous==2.2.0
+jedi==0.19.2
+Jinja2==3.1.6
+jiter==0.12.0
+joblib==1.5.2
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+kimina-client==0.2.1
+kiwisolver==1.4.9
+lark==1.2.2
+llguidance==0.7.30
+llvmlite==0.44.0
+lm-format-enforcer==0.10.12
+loguru==0.7.3
+Markdown==3.10
+markdown-it-py==4.0.0
+matplotlib==3.10.7
+matplotlib-inline==0.2.1
+mcp==1.22.0
+mdurl==0.1.2
+mistral_common==1.8.6
+mpmath==1.3.0
+msgpack==1.1.2
+msgspec==0.20.0
+multidict==6.7.0
+multiprocess==0.70.18
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+nest-asyncio==1.6.0
+networkx==3.3
+ninja==1.13.0
+nltk==3.9.2
+numba==0.61.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-cusparselt-cu12==0.6.2
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+omegaconf==2.3.0
+onnxruntime==1.23.2
+openai==2.8.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opencv-python-headless==4.12.0.88
+opentelemetry-api==1.38.0
+opentelemetry-exporter-otlp==1.26.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+opentelemetry-exporter-prometheus==0.59b0
+opentelemetry-proto==1.26.0
+opentelemetry-sdk==1.38.0
+opentelemetry-semantic-conventions==0.59b0
+opentelemetry-semantic-conventions-ai==0.4.13
+orjson==3.11.4
+outlines==0.1.11
+outlines_core==0.1.26
+packaging==25.0
+pandas==2.3.3
+parso==0.8.5
+partial-json-parser==0.2.1.1.post7
+pathspec==0.12.1
+peft==0.18.0
+pexpect==4.9.0
+pillow==11.3.0
+pip==25.3
+platformdirs==4.5.0
+preshed==3.0.12
+prometheus_client==0.23.1
+prometheus-fastapi-instrumentator==7.1.0
+prompt_toolkit==3.0.52
+propcache==0.4.1
+proto-plus==1.26.1
+protobuf==4.25.8
+psutil==7.1.3
+ptyprocess==0.7.0
+pure_eval==0.2.3
+py-cpuinfo==9.0.0
+py-spy==0.4.1
+pyarrow==22.0.0
+pyasn1==0.6.1
+pyasn1_modules==0.4.2
+pybind11==3.0.1
+pycountry==24.6.1
+pycparser==2.23
+pydantic==2.12.5
+pydantic_core==2.41.5
+pydantic-extra-types==2.10.6
+pydantic-settings==2.12.0
+pygame==2.6.1
+Pygments==2.19.2
+pyjnius==1.7.0
+pylatexenc==2.10
+pyparsing==3.2.5
+pyserini==1.2.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.1
+python-json-logger==4.0.0
+python-multipart==0.0.20
+pytokens==0.3.0
+pytz==2025.2
+PyYAML==6.0.3
+pyzmq==27.1.0
+ragen==0.1
+rank-bm25==0.2.2
+RapidFuzz==3.14.3
+ray==2.52.1
+referencing==0.37.0
+regex==2025.11.3
+requests==2.32.5
+rich==14.2.0
+rich-toolkit==0.17.0
+rignore==0.7.6
+rpds-py==0.30.0
+rsa==4.9.1
+safetensors==0.7.0
+scikit-learn==1.7.2
+scipy==1.15.3
+sentencepiece==0.2.1
+sentry-sdk==2.46.0
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.17.0
+smart_open==7.5.0
+smmap==5.0.2
+sniffio==1.3.1
+soupsieve==2.8
+spacy==3.8.11
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.2
+sse-starlette==3.0.3
+stack-data==0.6.3
+starlette==0.50.0
+sympy==1.13.1
+tabulate==0.9.0
+tenacity==9.1.2
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+tensordict==0.8.3
+thefuzz==0.22.1
+thinc==8.3.10
+threadpoolctl==3.6.0
+tiktoken==0.12.0
+together==1.5.31
+tokenizers==0.22.1
+tomli==2.3.0
+torch==2.6.0+cu124
+torchaudio==2.6.0
+torchdata==0.11.0
+torchvision==0.21.0
+tqdm==4.67.1
+traitlets==5.14.3
+transformers==4.57.3
+triton==3.2.0
+typer==0.19.2
+typer-slim==0.20.0
+typing_extensions==4.15.0
+typing-inspection==0.4.2
+tzdata==2025.2
+urllib3==2.5.0
+uvicorn==0.38.0
+uvloop==0.22.1
+verl==0.5.0.dev0
+virtualenv==20.35.4
+vllm==0.8.5
+wandb==0.23.0
+wasabi==1.1.3
+watchfiles==1.1.1
+wcwidth==0.2.14
+weasel==0.4.3
+websockets==15.0.1
+Werkzeug==3.1.4
+wheel==0.45.1
+wrapt==2.0.1
+xformers==0.0.29.post2
+xgrammar==0.1.18
+xxhash==3.6.0
+yarl==1.22.0
+zipp==3.23.0
+typeguard==4.4.4
+tyro==1.0.3
+verl==0.5.0.dev0
+autocommand==2.2.2
+backports.tarfile==1.2.0
+importlib_metadata==8.0.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+jaraco.context==5.3.0
+jaraco.functools==4.0.1
+jaraco.text==3.12.1
+more-itertools==10.3.0
+packaging==24.2
+platformdirs==4.2.2
+tomli==2.0.1
+typeguard==4.3.0
+typing_extensions==4.12.2
+wheel==0.45.1
+zipp==3.19.2

wandb/run-20260112_224942-jlcvlgih/files/wandb-metadata.json

@@ -0,0 +1,64 @@
+{
+  "os": "Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.19",
+  "startedAt": "2026-01-12T14:49:42.621345Z",
+  "args": [
+    "--track"
+  ],
+  "program": "/mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePath": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePathLocal": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "git": {
+    "remote": "https://github.com/mll-lab-nu/RAGEN.git",
+    "commit": "8d73639b99b38265453f898b8d6af7d4af50d56e"
+  },
+  "email": "haoyu-wa22@mails.tsinghua.edu.cn",
+  "root": "/mnt/general/wanghy/RAGEN",
+  "host": "pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0",
+  "executable": "/root/local/miniconda3/envs/ragen/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "2576980377600",
+      "used": "3724308480"
+    }
+  },
+  "memory": {
+    "total": "2163639439360"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-cff454b9-f118-535b-5e1f-6d503e78e52e"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9e77c178-188c-f869-780e-858c20a7d322"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "x6hdflogtnk75ni5ehl9wyx5sede6lv4"
+}

wandb/run-20260112_224942-jlcvlgih/files/wandb-summary.json

@@ -0,0 +1 @@
+{"charts/epsilon":0.05,"_timestamp":1.7682295877963042e+09,"eval/std_return":0.36697596651552,"perf/SPS":652,"_runtime":203.831261036,"eval/avg_return":-0.7192000000000002,"train/loss":0.01819211058318615,"train/episodic_length":20,"global_step":130108,"eval/episodes":4000,"_wandb":{"runtime":203},"train/episodic_return":-0.8000000000000002,"train/success_rate_100":0.01,"eval/success_rate":0.04625,"_step":130108,"train/success":0,"train/learning_rate":0.00025}

wandb/run-20260112_225250-k2i6dnry/files/code/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py

@@ -0,0 +1,485 @@
+# DQN with Action Masking for RAGEN Sudoku (4x4, max_step=20)
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.sudoku.env import SudokuEnv
+from ragen.env.sudoku.config import SudokuEnvConfig
+
+
+class SudokuWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: SudokuEnv, grid_size: int):
+        super().__init__()
+        self._env = env
+        self._size = grid_size
+        self._val_dim = self._size + 1
+        self._act_n = self._size * self._size * self._size
+        self.action_space = gym.spaces.Discrete(self._act_n)
+        self.observation_space = gym.spaces.Dict({
+            "observation": gym.spaces.Box(low=0.0, high=1.0, shape=(self._size * self._size * self._val_dim,), dtype=np.float32),
+            "action_mask": gym.spaces.Box(low=0.0, high=1.0, shape=(self._act_n,), dtype=np.float32)
+        })
+
+    def _encode_obs(self, text_obs: str) -> Dict[str, np.ndarray]:
+        vals: List[int] = []
+        for line in text_obs.splitlines():
+            ls = line.strip()
+            if len(ls) == 0:
+                continue
+            if set(ls) <= {'-'}:
+                continue
+            tokens = [t for t in ls.split() if t != '|']
+            if len(tokens) == 0:
+                continue
+            for t in tokens:
+                if t == '.':
+                    vals.append(0)
+                else:
+                    try:
+                        v = int(t)
+                    except ValueError:
+                        v = 0
+                    vals.append(v)
+        target = self._size * self._size
+        if len(vals) < target:
+            vals.extend([0] * (target - len(vals)))
+        if len(vals) > target:
+            vals = vals[:target]
+        grid = np.zeros((target, self._val_dim), dtype=np.float32)
+        mask = np.zeros(self._act_n, dtype=np.float32)
+        for i, v in enumerate(vals):
+            v_clamped = int(v)
+            if v_clamped < 0 or v_clamped > self._size:
+                v_clamped = 0
+            grid[i, v_clamped] = 1.0
+            if v_clamped != 0:
+                start_idx = i * self._size
+                end_idx = start_idx + self._size
+                mask[start_idx:end_idx] = 0.0
+        cur = np.array(vals, dtype=np.int64).reshape(self._size, self._size)
+        box = int(np.sqrt(self._size))
+        for i, v in enumerate(vals):
+            if int(v) == 0:
+                r = i // self._size
+                c = i % self._size
+                row_vals = set(cur[r, :].tolist())
+                col_vals = set(cur[:, c].tolist())
+                br = (r // box) * box
+                bc = (c // box) * box
+                box_vals = set(cur[br:br + box, bc:bc + box].reshape(-1).tolist())
+                start_idx = i * self._size
+                for num in range(1, self._size + 1):
+                    if (num not in row_vals) and (num not in col_vals) and (num not in box_vals):
+                        mask[start_idx + (num - 1)] = 1.0
+        if mask.sum() == 0:
+            for i, v in enumerate(vals):
+                if int(v) == 0:
+                    start_idx = i * self._size
+                    end_idx = start_idx + self._size
+                    mask[start_idx:end_idx] = 1.0
+            if mask.sum() == 0:
+                mask[:] = 1.0
+        return {
+            "observation": grid.reshape(-1),
+            "action_mask": mask,
+        }
+
+    @staticmethod
+    def _decode_action(action_id: int, grid_size: int) -> Tuple[int, int, int]:
+        g = grid_size
+        row = action_id // (g * g)
+        rem = action_id % (g * g)
+        col = rem // g
+        num = (rem % g) + 1
+        return row, col, num
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        row, col, num = self._decode_action(int(action), self._size)
+        act_str = f"{row + 1},{col + 1},{num}"
+        text_obs, reward, done, info = self._env.step(act_str)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        return obs, float(reward), terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "SudokuDQN"
+    total_timesteps: int = 200000
+    learning_rate: float = 3e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 400_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 1
+    learning_starts: int = 10000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Sudoku specific
+    grid_size: int = 4
+    difficulty: str = "easy"
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, grid_size, difficulty, capture_video=False):
+    def thunk():
+        config = SudokuEnvConfig(
+            grid_size=grid_size,
+            difficulty=difficulty,
+            render_mode='text',
+            render_format='simple',
+        )
+        env = SudokuEnv(config)
+        env = SudokuWrapper(env, grid_size)
+        max_steps = 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_dim: int):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity, obs_dim), dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity, obs_dim), dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.grid_size, args.difficulty, args.capture_video)()
+    # Flattened grid only from Dict observation
+    sample_obs, _ = env.reset(seed=args.seed)
+    obs_dim = int(np.prod(sample_obs["observation"].shape))
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(obs_dim, act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(obs_dim, act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_dim)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                obs_dict, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                state = obs_dict['observation']
+                mask = obs_dict['action_mask']
+                traj_states = [state.tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.grid_size * args.grid_size * 6)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        mask_t = torch.tensor(mask, dtype=torch.float32, device=device).unsqueeze(0)
+                        masked_q = q + (mask_t - 1.0) * 1e8
+                        action = int(torch.argmax(masked_q, dim=1).item())
+                    next_obs_dict, reward, terminated, truncated, info = env_eval.step(action)
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_obs_dict['observation']
+                    mask = next_obs_dict['action_mask']
+                    traj_states.append(state.tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    global_step = 0
+    start_time = time.time()
+
+    obs_dict, _ = env.reset(seed=args.seed)
+    obs = obs_dict['observation']
+    mask = obs_dict['action_mask']
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        with torch.no_grad():
+            q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+            mask_t = torch.tensor(mask, dtype=torch.float32, device=device).unsqueeze(0)
+            masked_q = q_values + (mask_t - 1.0) * 1e8
+            greedy_action = int(torch.argmax(masked_q, dim=1).item())
+        if (np.random.rand() < epsilon) or (global_step < args.learning_starts):
+            valid = np.where(mask > 0.5)[0]
+            if len(valid) > 0:
+                action = int(np.random.choice(valid))
+            else:
+                action = int(np.random.randint(0, act_dim))
+        else:
+            action = greedy_action
+
+        next_obs_dict, reward, terminated, truncated, info = env.step(action)
+        done = bool(terminated) or bool(truncated)
+
+        next_obs = next_obs_dict['observation']
+        next_mask = next_obs_dict['action_mask']
+
+        rb.add(obs.astype(np.float32), int(action), float(reward), bool(done), next_obs.astype(np.float32))
+
+        obs = next_obs
+        mask = next_mask
+        global_step += 1
+
+        if done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "rollout/success": float(1.0 if succ else 0.0),
+                        "rollout/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs_dict, _ = env.reset()
+            obs = obs_dict['observation']
+            mask = obs_dict['action_mask']
+
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    return make_env(0, run_name, args.seed + 9999, args.grid_size, args.difficulty, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_225250-k2i6dnry/files/config.yaml

@@ -0,0 +1,167 @@
+_wandb:
+    value:
+        cli_version: 0.23.0
+        code_path: code/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+        e:
+            42u9xwkov28g9dcp5599r8ywxjoe0be9:
+                args:
+                    - --track
+                codePath: cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+                codePathLocal: cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "2576980377600"
+                        used: "3724312576"
+                email: haoyu-wa22@mails.tsinghua.edu.cn
+                executable: /root/local/miniconda3/envs/ragen/bin/python
+                git:
+                    commit: 8d73639b99b38265453f898b8d6af7d4af50d56e
+                    remote: https://github.com/mll-lab-nu/RAGEN.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-cff454b9-f118-535b-5e1f-6d503e78e52e
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9e77c178-188c-f869-780e-858c20a7d322
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7
+                host: pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0
+                memory:
+                    total: "2163639439360"
+                os: Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35
+                program: /mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py
+                python: CPython 3.10.19
+                root: /mnt/general/wanghy/RAGEN
+                startedAt: "2026-01-12T14:52:50.232875Z"
+                writerId: 42u9xwkov28g9dcp5599r8ywxjoe0be9
+        m:
+            - "1": global_step
+              "6":
+                - 3
+              "7": []
+            - "2": charts/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": perf/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": train/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": rollout/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": eval/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": losses/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+        python_version: 3.10.19
+        t:
+            "1":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "2":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "3":
+                - 7
+                - 13
+                - 16
+                - 61
+            "4": 3.10.19
+            "5": 0.23.0
+            "6": 4.57.3
+            "12": 0.23.0
+            "13": linux-x86_64
+batch_size:
+    value: 128
+buffer_size:
+    value: 400000
+capture_video:
+    value: false
+cuda:
+    value: true
+difficulty:
+    value: easy
+end_e:
+    value: 0.05
+env_id:
+    value: SudokuDQN
+eval_episodes:
+    value: 4000
+eval_splits:
+    value: 2
+exp_name:
+    value: dqn_sudoku
+exploration_fraction:
+    value: 0.2
+gamma:
+    value: 0.99
+grid_size:
+    value: 4
+hidden_size:
+    value: 256
+learning_rate:
+    value: 0.0003
+learning_starts:
+    value: 10000
+seed:
+    value: 1
+start_e:
+    value: 1
+target_network_frequency:
+    value: 4000
+torch_deterministic:
+    value: true
+total_timesteps:
+    value: 200000
+track:
+    value: true
+train_frequency:
+    value: 1
+wandb_entity:
+    value: null
+wandb_project_name:
+    value: cleanRL

wandb/run-20260112_225250-k2i6dnry/files/diff.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_225250-k2i6dnry/files/diff_8d73639b99b38265453f898b8d6af7d4af50d56e.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_225250-k2i6dnry/files/requirements.txt

@@ -0,0 +1,313 @@
+colorama==0.4.6
+psutil==7.1.3
+ragen==0.1
+Farama-Notifications==0.0.4
+MarkupSafe==2.1.5
+PyJWT==2.10.1
+PySocks==1.7.1
+absl-py==2.3.1
+accelerate==1.12.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.2
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+airportsdata==20250909
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anthropic==0.75.0
+antlr4-python3-runtime==4.9.3
+anyio==4.12.0
+astor==0.8.1
+asttokens==3.0.1
+async-timeout==5.0.1
+attrs==25.4.0
+beautifulsoup4==4.14.3
+black==25.11.0
+blake3==1.0.8
+blinker==1.9.0
+blis==1.3.3
+cachetools==6.2.2
+catalogue==2.0.10
+certifi==2025.11.12
+cffi==2.0.0
+charset-normalizer==3.4.4
+cleantext==1.1.4
+click==8.2.1
+cloudpathlib==0.23.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+coloredlogs==15.0.1
+colorful==0.5.8
+compressed-tensors==0.9.3
+confection==0.1.5
+contourpy==1.3.2
+cryptography==46.0.3
+cupy-cuda12x==13.6.0
+cycler==0.12.1
+cymem==2.0.13
+Cython==3.2.2
+datasets==4.4.1
+debugpy==1.8.17
+decorator==5.2.1
+Deprecated==1.3.1
+depyf==0.18.0
+dill==0.4.0
+diskcache==5.6.3
+distlib==0.4.0
+distro==1.9.0
+dnspython==2.8.0
+docstring_parser==0.17.0
+einops==0.8.1
+email-validator==2.3.0
+eval_type_backport==0.2.2
+exceptiongroup==1.3.1
+executing==2.2.1
+faiss-gpu==1.7.2
+fastapi==0.123.1
+fastapi-cli==0.0.16
+fastapi-cloud-cli==0.5.2
+fastar==0.8.0
+fastrlock==0.8.3
+filelock==3.19.1
+flash_attn==2.8.3
+Flask==3.1.2
+flatbuffers==25.9.23
+fonttools==4.61.0
+frozenlist==1.8.0
+fsspec==2025.9.0
+gdown==5.2.0
+gguf==0.17.1
+gitdb==4.0.12
+GitPython==3.1.45
+google-api-core==2.28.1
+google-auth==2.43.0
+googleapis-common-protos==1.72.0
+grpcio==1.76.0
+gym==0.26.2
+gym-notices==0.1.0
+gym-sokoban==0.0.6
+gymnasium==1.2.2
+h11==0.16.0
+hf-xet==1.2.0
+html2text==2025.4.15
+httpcore==1.0.9
+httptools==0.7.1
+httpx==0.28.1
+httpx-sse==0.4.3
+huggingface-hub==0.36.0
+humanfriendly==10.0
+hydra-core==1.3.2
+idna==3.11
+ImageIO==2.37.2
+importlib_metadata==8.0.0
+interegular==0.3.3
+ipython==8.37.0
+itsdangerous==2.2.0
+jedi==0.19.2
+Jinja2==3.1.6
+jiter==0.12.0
+joblib==1.5.2
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+kimina-client==0.2.1
+kiwisolver==1.4.9
+lark==1.2.2
+llguidance==0.7.30
+llvmlite==0.44.0
+lm-format-enforcer==0.10.12
+loguru==0.7.3
+Markdown==3.10
+markdown-it-py==4.0.0
+matplotlib==3.10.7
+matplotlib-inline==0.2.1
+mcp==1.22.0
+mdurl==0.1.2
+mistral_common==1.8.6
+mpmath==1.3.0
+msgpack==1.1.2
+msgspec==0.20.0
+multidict==6.7.0
+multiprocess==0.70.18
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+nest-asyncio==1.6.0
+networkx==3.3
+ninja==1.13.0
+nltk==3.9.2
+numba==0.61.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-cusparselt-cu12==0.6.2
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+omegaconf==2.3.0
+onnxruntime==1.23.2
+openai==2.8.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opencv-python-headless==4.12.0.88
+opentelemetry-api==1.38.0
+opentelemetry-exporter-otlp==1.26.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+opentelemetry-exporter-prometheus==0.59b0
+opentelemetry-proto==1.26.0
+opentelemetry-sdk==1.38.0
+opentelemetry-semantic-conventions==0.59b0
+opentelemetry-semantic-conventions-ai==0.4.13
+orjson==3.11.4
+outlines==0.1.11
+outlines_core==0.1.26
+packaging==25.0
+pandas==2.3.3
+parso==0.8.5
+partial-json-parser==0.2.1.1.post7
+pathspec==0.12.1
+peft==0.18.0
+pexpect==4.9.0
+pillow==11.3.0
+pip==25.3
+platformdirs==4.5.0
+preshed==3.0.12
+prometheus_client==0.23.1
+prometheus-fastapi-instrumentator==7.1.0
+prompt_toolkit==3.0.52
+propcache==0.4.1
+proto-plus==1.26.1
+protobuf==4.25.8
+psutil==7.1.3
+ptyprocess==0.7.0
+pure_eval==0.2.3
+py-cpuinfo==9.0.0
+py-spy==0.4.1
+pyarrow==22.0.0
+pyasn1==0.6.1
+pyasn1_modules==0.4.2
+pybind11==3.0.1
+pycountry==24.6.1
+pycparser==2.23
+pydantic==2.12.5
+pydantic_core==2.41.5
+pydantic-extra-types==2.10.6
+pydantic-settings==2.12.0
+pygame==2.6.1
+Pygments==2.19.2
+pyjnius==1.7.0
+pylatexenc==2.10
+pyparsing==3.2.5
+pyserini==1.2.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.1
+python-json-logger==4.0.0
+python-multipart==0.0.20
+pytokens==0.3.0
+pytz==2025.2
+PyYAML==6.0.3
+pyzmq==27.1.0
+ragen==0.1
+rank-bm25==0.2.2
+RapidFuzz==3.14.3
+ray==2.52.1
+referencing==0.37.0
+regex==2025.11.3
+requests==2.32.5
+rich==14.2.0
+rich-toolkit==0.17.0
+rignore==0.7.6
+rpds-py==0.30.0
+rsa==4.9.1
+safetensors==0.7.0
+scikit-learn==1.7.2
+scipy==1.15.3
+sentencepiece==0.2.1
+sentry-sdk==2.46.0
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.17.0
+smart_open==7.5.0
+smmap==5.0.2
+sniffio==1.3.1
+soupsieve==2.8
+spacy==3.8.11
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.2
+sse-starlette==3.0.3
+stack-data==0.6.3
+starlette==0.50.0
+sympy==1.13.1
+tabulate==0.9.0
+tenacity==9.1.2
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+tensordict==0.8.3
+thefuzz==0.22.1
+thinc==8.3.10
+threadpoolctl==3.6.0
+tiktoken==0.12.0
+together==1.5.31
+tokenizers==0.22.1
+tomli==2.3.0
+torch==2.6.0+cu124
+torchaudio==2.6.0
+torchdata==0.11.0
+torchvision==0.21.0
+tqdm==4.67.1
+traitlets==5.14.3
+transformers==4.57.3
+triton==3.2.0
+typer==0.19.2
+typer-slim==0.20.0
+typing_extensions==4.15.0
+typing-inspection==0.4.2
+tzdata==2025.2
+urllib3==2.5.0
+uvicorn==0.38.0
+uvloop==0.22.1
+verl==0.5.0.dev0
+virtualenv==20.35.4
+vllm==0.8.5
+wandb==0.23.0
+wasabi==1.1.3
+watchfiles==1.1.1
+wcwidth==0.2.14
+weasel==0.4.3
+websockets==15.0.1
+Werkzeug==3.1.4
+wheel==0.45.1
+wrapt==2.0.1
+xformers==0.0.29.post2
+xgrammar==0.1.18
+xxhash==3.6.0
+yarl==1.22.0
+zipp==3.23.0
+typeguard==4.4.4
+tyro==1.0.3
+verl==0.5.0.dev0
+autocommand==2.2.2
+backports.tarfile==1.2.0
+importlib_metadata==8.0.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+jaraco.context==5.3.0
+jaraco.functools==4.0.1
+jaraco.text==3.12.1
+more-itertools==10.3.0
+packaging==24.2
+platformdirs==4.2.2
+tomli==2.0.1
+typeguard==4.3.0
+typing_extensions==4.12.2
+wheel==0.45.1
+zipp==3.19.2

wandb/run-20260112_225250-k2i6dnry/files/wandb-metadata.json

@@ -0,0 +1,64 @@
+{
+  "os": "Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.19",
+  "startedAt": "2026-01-12T14:52:50.232875Z",
+  "args": [
+    "--track"
+  ],
+  "program": "/mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_sudoku.py",
+  "codePath": "cleanrl/cleanrl/scout_dqn/dqn_sudoku.py",
+  "codePathLocal": "cleanrl/cleanrl/scout_dqn/dqn_sudoku.py",
+  "git": {
+    "remote": "https://github.com/mll-lab-nu/RAGEN.git",
+    "commit": "8d73639b99b38265453f898b8d6af7d4af50d56e"
+  },
+  "email": "haoyu-wa22@mails.tsinghua.edu.cn",
+  "root": "/mnt/general/wanghy/RAGEN",
+  "host": "pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0",
+  "executable": "/root/local/miniconda3/envs/ragen/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "2576980377600",
+      "used": "3724312576"
+    }
+  },
+  "memory": {
+    "total": "2163639439360"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-cff454b9-f118-535b-5e1f-6d503e78e52e"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9e77c178-188c-f869-780e-858c20a7d322"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "42u9xwkov28g9dcp5599r8ywxjoe0be9"
+}

wandb/run-20260112_225250-k2i6dnry/files/wandb-summary.json

@@ -0,0 +1 @@
+{"eval/success_rate":0.76425,"rollout/success_rate_100":0.71,"train/loss":2959.95068359375,"eval/episodes":4000,"charts/epsilon":0.05,"eval/std_return":6.09278240523162,"_timestamp":1.7682304639431205e+09,"rollout/success":1,"perf/SPS":230,"global_step":200000,"_step":200000,"eval/avg_return":12.58475,"_runtime":892.284096128,"_wandb":{"runtime":892},"train/learning_rate":0.0003}

wandb/run-20260112_225324-locrg3kr/files/code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py

@@ -0,0 +1,454 @@
+# DQN with Step Penalty & Time Limit Bootstrap for RAGEN Rubik's Cube 2x2
+import os
+import random
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple, Dict, Any, List
+import json
+import re
+
+import gymnasium as gym
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import tyro
+
+import sys
+# 假设 ragen 库在当前目录的上两级，请根据实际情况调整
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../../'))
+
+from ragen.env.rubikscube.env import RubiksCube2x2Env
+from ragen.env.rubikscube.config import RubiksCube2x2Config
+
+
+class RubiksCubeWrapper(gym.Env):
+    metadata = {"render_modes": ["rgb_array", "human", "ansi"]}
+
+    def __init__(self, env: RubiksCube2x2Env):
+        super().__init__()
+        self._env = env
+        self._colors = ['W', 'O', 'G', 'R', 'B', 'Y']
+        self._color_to_idx = {c: i for i, c in enumerate(self._colors)}
+        self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(24 * len(self._colors),), dtype=np.float32)
+        self.action_space = gym.spaces.Discrete(12)
+        self._face_pat = re.compile(r"\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])\n\s*\[(?:\s*([A-Z])\s*,\s*([A-Z])\s*\])")
+        
+        # [关键修改] 每步惩罚
+        # 如果20步都没解出来，累计惩罚是 20 * -0.05 = -1.0，抵消掉最后可能获得的 +1.0
+        # 这迫使 agent 寻找更短路径
+        self.step_penalty = -0.044
+
+    def _encode_obs(self, text_obs: str) -> np.ndarray:
+        faces_order = ["Up (U):", "Left (L):", "Front (F):", "Right (R):", "Back (B):", "Down (D):"]
+        lines = text_obs.splitlines()
+        blocks: List[str] = []
+        i = 0
+        while i < len(lines):
+            line = lines[i]
+            for header in faces_order:
+                if line.startswith(header):
+                    content = line[len(header):].strip()
+                    next_line = lines[i + 1] if i + 1 < len(lines) else ""
+                    block = f"{content}\n{next_line}"
+                    blocks.append(block)
+                    break
+            i += 1
+        if len(blocks) != 6:
+            return np.zeros(24 * len(self._colors), dtype=np.float32)
+        stickers: List[int] = []
+        for blk in blocks:
+            m = self._face_pat.search(blk)
+            if not m:
+                return np.zeros(24 * len(self._colors), dtype=np.float32)
+            c0, c1, c2, c3 = m.group(1), m.group(2), m.group(3), m.group(4)
+            stickers.extend([c0, c1, c2, c3])
+        grid = np.zeros((24, len(self._colors)), dtype=np.float32)
+        for idx, ch in enumerate(stickers):
+            cidx = self._color_to_idx.get(ch, None)
+            if cidx is not None:
+                grid[idx, cidx] = 1.0
+        return grid.reshape(-1)
+
+    def reset(self, *, seed: int | None = None, options: Dict[str, Any] | None = None):
+        text_obs = self._env.reset(seed=seed)
+        obs = self._encode_obs(text_obs)
+        return obs, {}
+
+    def step(self, action: int):
+        mapped = int(action) + 1
+        text_obs, reward, done, info = self._env.step(mapped)
+        obs = self._encode_obs(text_obs)
+        terminated = bool(done)
+        truncated = False
+        
+        # [关键修改] 应用惩罚
+        # 此时 reward 包含原本的稀疏奖励 (0 或 1) 加上每步惩罚
+        reward = float(reward) + self.step_penalty
+        
+        return obs, reward, terminated, truncated, info or {}
+
+    def render(self):
+        return self._env.render()
+
+    def close(self):
+        self._env.close()
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    seed: int = 1
+    torch_deterministic: bool = True
+    cuda: bool = True
+    track: bool = True
+    wandb_project_name: str = "cleanRL"
+    wandb_entity: str | None = None
+    capture_video: bool = False
+
+    # Algorithm
+    env_id: str = "RubiksCube2x2DQN"
+    total_timesteps: int = 400_000
+    learning_rate: float = 2.5e-4
+    gamma: float = 0.99
+    batch_size: int = 128
+    buffer_size: int = 200_000
+    target_network_frequency: int = 4000
+    train_frequency: int = 4
+    learning_starts: int = 5000
+
+    # Epsilon-greedy
+    start_e: float = 1.0
+    end_e: float = 0.05
+    exploration_fraction: float = 0.2
+
+    # Model size
+    hidden_size: int = 256
+
+    # Rubik specific
+    scramble_depth: int = 3 # 初始难度
+    max_steps_env: int = 20 # 硬性时间限制
+
+    # Eval
+    eval_splits: int = 2
+    eval_episodes: int = 4000
+
+
+def make_env(idx, run_name, seed, scramble_depth, max_steps_env, capture_video=False):
+    def thunk():
+        # [关键修改] 内部 max_steps 设为 1000，防止 env 内部发出 done=True
+        # 我们完全依赖外部 TimeLimit wrapper 来处理超时
+        config = RubiksCube2x2Config(scramble_depth=scramble_depth, max_steps=20, render_mode='text')
+        env = RubiksCube2x2Env(config)
+        env = RubiksCubeWrapper(env)
+        # 外部限制设为 20
+        env = gym.wrappers.TimeLimit(env, max_episode_steps=max_steps_env)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video and idx == 0:
+            env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        return env
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class QNetwork(nn.Module):
+    def __init__(self, obs_dim: int, act_dim: int, hidden: int):
+        super().__init__()
+        self.net = nn.Sequential(
+            layer_init(nn.Linear(obs_dim, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, hidden)),
+            nn.ReLU(),
+            layer_init(nn.Linear(hidden, act_dim), std=0.01),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class ReplayBuffer:
+    def __init__(self, capacity: int, obs_shape: Tuple[int, ...]):
+        self.capacity = capacity
+        self.ptr = 0
+        self.full = False
+        self.obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.next_obs_buf = np.zeros((capacity,) + obs_shape, dtype=np.float32)
+        self.act_buf = np.zeros((capacity,), dtype=np.int64)
+        self.rew_buf = np.zeros((capacity,), dtype=np.float32)
+        self.done_buf = np.zeros((capacity,), dtype=np.float32)
+
+    def add(self, obs: np.ndarray, act: int, rew: float, done: bool, next_obs: np.ndarray):
+        i = self.ptr
+        self.obs_buf[i] = obs
+        self.next_obs_buf[i] = next_obs
+        self.act_buf[i] = act
+        self.rew_buf[i] = rew
+        self.done_buf[i] = 1.0 if done else 0.0
+        self.ptr = (self.ptr + 1) % self.capacity
+        if self.ptr == 0:
+            self.full = True
+
+    def can_sample(self, batch_size: int) -> bool:
+        return (self.capacity if self.full else self.ptr) >= batch_size
+
+    def sample(self, batch_size: int):
+        size = self.capacity if self.full else self.ptr
+        idxs = np.random.randint(0, size, size=batch_size)
+        return (
+            self.obs_buf[idxs],
+            self.act_buf[idxs],
+            self.rew_buf[idxs],
+            self.done_buf[idxs],
+            self.next_obs_buf[idxs],
+        )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(Args)
+    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
+
+    if args.track:
+        import wandb
+        wandb.init(
+            project=args.wandb_project_name,
+            entity=args.wandb_entity,
+            config=vars(args),
+            name=run_name,
+            monitor_gym=True,
+            save_code=True,
+        )
+        try:
+            wandb.define_metric("global_step")
+            for prefix in ["train/*", "rollout/*", "eval/*", "losses/*", "charts/*", "perf/*"]:
+                wandb.define_metric(prefix, step_metric="global_step")
+        except Exception:
+            pass
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.backends.cudnn.deterministic = args.torch_deterministic
+
+    device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
+
+    env = make_env(0, run_name, args.seed, args.scramble_depth, args.max_steps_env, args.capture_video)()
+    obs_shape = env.observation_space.shape
+    act_dim = env.action_space.n
+
+    policy_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net = QNetwork(int(np.prod(obs_shape)), act_dim, args.hidden_size).to(device)
+    target_net.load_state_dict(policy_net.state_dict())
+    target_net.eval()
+
+    optimizer = optim.Adam(policy_net.parameters(), lr=args.learning_rate)
+    criterion = nn.SmoothL1Loss()
+
+    rb = ReplayBuffer(args.buffer_size, obs_shape)
+
+    exploration_steps = max(1, int(args.exploration_fraction * args.total_timesteps))
+    epsilon_by_step = lambda t: args.end_e + (args.start_e - args.end_e) * max(0.0, (exploration_steps - t) / exploration_steps)
+
+    # 评估函数
+    def collect_eval_trajectories(agent_model, make_env_fn, n_episodes, step_tag):
+        out_dir = Path(f"runs/{run_name}/trajectories/step_{step_tag}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+        out_path = out_dir / "trajectories.jsonl"
+        env_eval = make_env_fn()
+        collected = 0
+        summary_returns = []
+        summary_success = []
+        with out_path.open("w") as f:
+            while collected < n_episodes:
+                state, _ = env_eval.reset(seed=args.seed + 100000 + collected)
+                traj_states = [np.asarray(state).tolist()]
+                traj_actions = []
+                traj_rewards = []
+                traj_dones = []
+                traj_success = []
+                done = False
+                step_count = 0
+                max_eval_steps = getattr(env_eval, '_max_episode_steps', None) or int(args.max_steps_env)
+                while not done:
+                    with torch.no_grad():
+                        q = agent_model(torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0))
+                        action = int(torch.argmax(q, dim=1).item())
+                    next_state, reward, terminated, truncated, info = env_eval.step(action)
+                    
+                    traj_actions.append(int(action))
+                    traj_rewards.append(float(reward))
+                    step_count += 1
+                    d = bool(terminated) or bool(truncated) or (step_count >= max_eval_steps)
+                    traj_dones.append(d)
+                    traj_success.append(bool((info or {}).get('success', False)))
+                    state = next_state
+                    traj_states.append(np.asarray(state).tolist())
+                    done = d
+                ep_ret = float(sum(traj_rewards))
+                ep_succ = bool(any(traj_success))
+                record = {
+                    "states": traj_states,
+                    "actions": traj_actions,
+                    "rewards": traj_rewards,
+                    "dones": traj_dones,
+                    "success": traj_success,
+                    "episode_return": ep_ret,
+                    "episode_success": ep_succ,
+                }
+                f.write(json.dumps(record) + "\n")
+                collected += 1
+                summary_returns.append(ep_ret)
+                summary_success.append(1.0 if ep_succ else 0.0)
+        env_eval.close()
+        try:
+            metrics = {
+                "global_step": int(step_tag),
+                "episodes": int(n_episodes),
+                "success_rate": float(np.mean(summary_success)) if len(summary_success) else 0.0,
+                "avg_return": float(np.mean(summary_returns)) if len(summary_returns) else 0.0,
+                "std_return": float(np.std(summary_returns)) if len(summary_returns) else 0.0,
+            }
+            with (out_dir / "metrics.json").open("w") as mf:
+                json.dump(metrics, mf)
+        except Exception as e:
+            print(f"Warning: failed to write eval metrics: {e}")
+
+    # 主训练循环
+    global_step = 0
+    start_time = time.time()
+
+    obs, _ = env.reset(seed=args.seed)
+    ep_return = 0.0
+    ep_len = 0
+    ep_success_window: List[float] = []
+
+    eval_every_steps = max(1, args.total_timesteps // args.eval_splits)
+
+    while global_step < args.total_timesteps:
+        epsilon = epsilon_by_step(global_step)
+        if np.random.rand() < epsilon or global_step < args.learning_starts:
+            action = env.action_space.sample()
+        else:
+            with torch.no_grad():
+                q_values = policy_net(torch.tensor(obs, dtype=torch.float32, device=device).unsqueeze(0))
+                action = int(torch.argmax(q_values, dim=1).item())
+
+        next_obs, reward, terminated, truncated, info = env.step(action)
+        
+        # 判断是否需要 Reset (任何结束都 Reset)
+        real_done = bool(terminated) or bool(truncated)
+
+        # [核心逻辑]
+        # 1. 即使是 truncated (超时)，done 也记为 False，以便进行 bootstrap (计算未来价值)
+        # 2. 只有 terminated (真正解开了)，done 才记为 True
+        # 3. Step Penalty 已经包含在 reward 中，DQN 会学到"为了避免扣分，必须在未来几步内解决"
+        rb.add(
+            obs.astype(np.float32), 
+            int(action), 
+            float(reward), 
+            bool(terminated), # 注意：只用 terminated
+            next_obs.astype(np.float32)
+        )
+
+        obs = next_obs
+        ep_return += float(reward)
+        ep_len += 1
+        global_step += 1
+
+        # 训练过程
+        if rb.can_sample(args.batch_size) and (global_step % args.train_frequency == 0) and (global_step > args.learning_starts):
+            batch_obs, batch_act, batch_rew, batch_done, batch_next_obs = rb.sample(args.batch_size)
+            b_obs = torch.tensor(batch_obs, dtype=torch.float32, device=device)
+            b_act = torch.tensor(batch_act, dtype=torch.int64, device=device)
+            b_rew = torch.tensor(batch_rew, dtype=torch.float32, device=device)
+            b_done = torch.tensor(batch_done, dtype=torch.float32, device=device)
+            b_next_obs = torch.tensor(batch_next_obs, dtype=torch.float32, device=device)
+
+            with torch.no_grad():
+                next_q = target_net(b_next_obs).max(dim=1)[0]
+                target_q = b_rew + args.gamma * (1.0 - b_done) * next_q
+
+            current_q = policy_net(b_obs).gather(1, b_act.view(-1, 1)).squeeze(1)
+            loss = criterion(current_q, target_q)
+
+            optimizer.zero_grad()
+            loss.backward()
+            nn.utils.clip_grad_norm_(policy_net.parameters(), max_norm=10.0)
+            optimizer.step()
+
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/loss": float(loss.item()),
+                        "charts/epsilon": float(epsilon),
+                        "perf/SPS": int(global_step / (time.time() - start_time)),
+                        "train/learning_rate": float(optimizer.param_groups[0]["lr"]),
+                    }, step=global_step)
+                except Exception:
+                    pass
+
+        # 结束时 Reset
+        if real_done:
+            succ = bool((info or {}).get('success', False))
+            ep_success_window.append(1.0 if succ else 0.0)
+            if len(ep_success_window) > 100:
+                ep_success_window.pop(0)
+            if args.track:
+                try:
+                    import wandb
+                    wandb.log({
+                        "global_step": int(global_step),
+                        "train/episodic_return": float(ep_return),
+                        "train/episodic_length": int(ep_len),
+                        "train/success": float(1.0 if succ else 0.0),
+                        "train/success_rate_100": float(np.mean(ep_success_window)) if len(ep_success_window) >= 1 else None,
+                    }, step=global_step)
+                except Exception:
+                    pass
+            obs, _ = env.reset()
+            ep_return, ep_len = 0.0, 0
+
+        # 更新目标网络
+        if global_step % args.target_network_frequency == 0:
+            target_net.load_state_dict(policy_net.state_dict())
+
+        # 打印日志
+        if global_step % 1000 == 0:
+            sps = int(global_step / (time.time() - start_time))
+            sr100 = float(np.mean(ep_success_window)) if len(ep_success_window) > 0 else 0.0
+            print(f"Step {global_step} | SPS: {sps} | Epsilon: {epsilon:.3f} | SR@100: {sr100:.3f}")
+
+        # 评估
+        if global_step == 1 or (global_step % eval_every_steps == 0):
+            try:
+                def eval_thunk():
+                    # 评估环境也同样设置：内部1000，外部20
+                    return make_env(0, run_name, args.seed + 9999, args.scramble_depth, args.max_steps_env, False)()
+                collect_eval_trajectories(policy_net, eval_thunk, n_episodes=args.eval_episodes, step_tag=global_step)
+                if args.track:
+                    try:
+                        import wandb
+                        mpath = Path(f"runs/{run_name}/trajectories/step_{global_step}/metrics.json")
+                        if mpath.exists():
+                            with mpath.open("r") as mf:
+                                metrics = json.load(mf)
+                            wandb.log({
+                                "eval/success_rate": metrics.get("success_rate"),
+                                "eval/avg_return": metrics.get("avg_return"),
+                                "eval/std_return": metrics.get("std_return"),
+                                "eval/episodes": metrics.get("episodes"),
+                            }, step=global_step)
+                    except Exception:
+                        pass
+                print(f"Collected {args.eval_episodes} eval trajectories at step {global_step}")
+            except Exception as e:
+                print(f"Warning: eval trajectory collection failed at step {global_step}: {e}")
+
+    env.close()

wandb/run-20260112_225324-locrg3kr/files/config.yaml

@@ -0,0 +1,167 @@
+_wandb:
+    value:
+        cli_version: 0.23.0
+        code_path: code/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+        e:
+            38xx0mqyxzx0knb11tf2de7k74fbhknb:
+                args:
+                    - --track
+                codePath: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                codePathLocal: cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "2576980377600"
+                        used: "3724316672"
+                email: haoyu-wa22@mails.tsinghua.edu.cn
+                executable: /root/local/miniconda3/envs/ragen/bin/python
+                git:
+                    commit: 8d73639b99b38265453f898b8d6af7d4af50d56e
+                    remote: https://github.com/mll-lab-nu/RAGEN.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-cff454b9-f118-535b-5e1f-6d503e78e52e
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9e77c178-188c-f869-780e-858c20a7d322
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7
+                host: pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0
+                memory:
+                    total: "2163639439360"
+                os: Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35
+                program: /mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py
+                python: CPython 3.10.19
+                root: /mnt/general/wanghy/RAGEN
+                startedAt: "2026-01-12T14:53:24.482263Z"
+                writerId: 38xx0mqyxzx0knb11tf2de7k74fbhknb
+        m:
+            - "1": global_step
+              "6":
+                - 3
+              "7": []
+            - "2": eval/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": losses/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": charts/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": perf/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": train/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+            - "2": rollout/*
+              "5": 1
+              "6":
+                - 1
+              "7": []
+        python_version: 3.10.19
+        t:
+            "1":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "2":
+                - 1
+                - 11
+                - 30
+                - 49
+                - 50
+                - 51
+                - 105
+            "3":
+                - 7
+                - 13
+                - 16
+                - 61
+            "4": 3.10.19
+            "5": 0.23.0
+            "6": 4.57.3
+            "12": 0.23.0
+            "13": linux-x86_64
+batch_size:
+    value: 128
+buffer_size:
+    value: 200000
+capture_video:
+    value: false
+cuda:
+    value: true
+end_e:
+    value: 0.05
+env_id:
+    value: RubiksCube2x2DQN
+eval_episodes:
+    value: 4000
+eval_splits:
+    value: 2
+exp_name:
+    value: dqn_rubikscube
+exploration_fraction:
+    value: 0.2
+gamma:
+    value: 0.99
+hidden_size:
+    value: 256
+learning_rate:
+    value: 0.00025
+learning_starts:
+    value: 5000
+max_steps_env:
+    value: 20
+scramble_depth:
+    value: 3
+seed:
+    value: 1
+start_e:
+    value: 1
+target_network_frequency:
+    value: 4000
+torch_deterministic:
+    value: true
+total_timesteps:
+    value: 400000
+track:
+    value: true
+train_frequency:
+    value: 4
+wandb_entity:
+    value: null
+wandb_project_name:
+    value: cleanRL

wandb/run-20260112_225324-locrg3kr/files/diff.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_225324-locrg3kr/files/diff_8d73639b99b38265453f898b8d6af7d4af50d56e.patch

@@ -0,0 +1,1248 @@
+diff --git a/config/base.yaml b/config/base.yaml
+index a1659e4..6029703 100644
+--- a/config/base.yaml
++++ b/config/base.yaml
+@@ -9,15 +9,16 @@ seed:
+   train: 10000
+   val: 123
+ 
+-micro_batch_size_per_gpu: 4
++micro_batch_size_per_gpu: 1
+ ppo_mini_batch_size: 32
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: 
++# /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+ enable_response_mask: True # Enabling response mask could improve stability of rollout/old_log_prob, as P(st|history) are no longer calculated in loss here. See https://docs.google.com/document/d/1bg7obeiKTExuHHBl5uOiSpec5uLDZ2Tgvxy6li5pHX4/edit?usp=sharing for more details.
+ grpo_advantage_length_weight: False # if you do not enable this and critic/advantage_estimator is GRPO, and the critic/advantages/mean is too low, then you can try enabling this to encourage reasoning and forbid collapse
+ 
+ lora:
+   rank: 0
+-  alpha: 64
++  alpha: 16
+   target_modules: all-linear
+ 
+ actor_rollout_ref:
+@@ -26,6 +27,7 @@ actor_rollout_ref:
+     lora_rank: ${lora.rank}
+     lora_alpha: ${lora.alpha}
+     target_modules: ${lora.target_modules}
++    torch_dtype: bfloat16
+   actor:
+     ppo_mini_batch_size: ${ppo_mini_batch_size}  # by default, ppo_mini_batch_size = train_batch_size / 4
+     ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+@@ -46,11 +48,11 @@ actor_rollout_ref:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu} # following micro_batch_size_per_gpu
+     tensor_model_parallel_size: 1
+-    max_model_len: 3600
++    max_model_len: 16384 #3600 why** 14400
+     prompt_length: 1 # useless. Just put it here
+     response_length: 400 # single-turn response length
+-    gpu_memory_utilization: 0.5
+-    max_num_batched_tokens: 8192 # set only when enable_chunked_prefill is true
++    gpu_memory_utilization: 0.7
++    max_num_batched_tokens: 16384 # set only when enable_chunked_prefill is true
+     temperature: 1
+     rollout_filter_ratio: 0.25
+     rollout_filter_type: largest # smallest or largest
+@@ -90,26 +92,28 @@ algorithm:
+     kl_coef: 0.000
+ 
+ trainer:
+-  project_name: ragen_latest
++  project_name: 
+   experiment_name: test
+   local_log_dir: "results/"
+-  save_freq: 100
++  save_freq: -1
+   total_training_steps: 200
+   validation_steps: 1 # validation instances = validation_steps * val_env_groups * group_size
+   val_before_train: True
+   n_gpus_per_node: 1
+   test_freq: 10
+   generations_to_log_to_wandb: 
++    train: 128
+     val: 20
+   logger: [ 'console', 'wandb' ]
+   max_actor_ckpt_to_keep: 1
+   max_critic_ckpt_to_keep: 1
++  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+ 
+ agent_proxy:
+   max_context_window: -1 # set a value > 0 to enable context window for long trajectory
+-  max_turn: 5
++  max_turn: 25 #25 why** 700
+   action_sep: "||"
+-  max_actions_per_turn: 2 # how many actions can be output at most in a single turn
++  max_actions_per_turn: 1 # how many actions can be output at most in a single turn
+   use_turn_scores: False # important to GAE when applying token-level rewards to token-level advantages. If False, will take the sum of scores as the reward for the last turn.
+   enable_think: True # False -> no think RL
+   reward_normalization:
+@@ -126,11 +130,11 @@ es_manager:
+       tags: ["CoordSokoban"]
+       n_groups: [8] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+   val:
+-    env_groups: 32
+-    group_size: 16 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
++    env_groups: 256
++    group_size: 1 # should be set to 1 because when val temperature is set to 0 and group size > 1, there will be repetitive prompts which leads to same trajectory.
+     env_configs:
+       tags: ["CoordSokoban"]
+-      n_groups: [32] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      n_groups: [256] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ ctx_manager:
+   generation: # go to vllm
+diff --git a/config/envs.yaml b/config/envs.yaml
+index 1002ac8..d258d15 100644
+--- a/config/envs.yaml
++++ b/config/envs.yaml
+@@ -2,6 +2,7 @@ custom_envs:
+   SimpleSokoban:
+     env_type: sokoban
+     max_actions_per_traj: 10 # used in environment state manager to control the actual max actions executed per trajectory
++    # ORIGNAL env_instruction
+     env_instruction: |
+       You are solving the Sokoban puzzle. 
+       You are the player and you need to push all boxes to targets. 
+@@ -12,7 +13,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+ 
+   LargerSokoban:
+@@ -65,7 +66,7 @@ custom_envs:
+     env_config: # keys should be a subset of SokobanConfig
+       dim_x: 6
+       dim_y: 6
+-      num_boxes: 1
++      num_boxes: 2
+       max_steps: 100
+       observation_format: "grid_coord"
+ 
+@@ -111,14 +112,14 @@ custom_envs:
+ 
+   FrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: "You are solving the FrozenLake puzzle. Forbid the whole and go to the target. You may move to the unintended direction due to the slippery ice. Example answer format: <think>To forbid the hole and go to the target, I should go left then go up.</think><answer>Left || Up</answer>"
+     max_tokens: 100
+     env_config: null
+ 
+   CoordFrozenLake:
+     env_type: frozen_lake
+-    max_actions_per_traj: 10
++    max_actions_per_traj: 25
+     env_instruction: |
+       You are solving the FrozenLake puzzle. The observation includes both a symbol grid and zero-indexed coordinates for the start, goal, player, and any holes.
+       Coordinates range from the top-left corner (0, 0) to the bottom-right corner (5, 5).
+@@ -195,4 +196,80 @@ custom_envs:
+     max_actions_per_traj: 30
+     env_instruction: "You are a Lean theorem prover. Given a Lean theorem statement, propose a sequence of tactics that completes the proof. Think step by step about which tactics to apply next. Provide tactics separated by '||', for example <answer>intro || simp || rfl</answer>."
+     max_tokens: 512
+-    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
+
++    env_config: null # Please refer to ragen/env/lean/config.py for a full list of parameters.
++
++  
++  game_2048:
++    env_type: game_2048
++    max_actions_per_traj: 700
++    env_instruction: |
++      You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.
++      If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.
++      Example: <answer>Up</answer>
++    max_tokens: 8192
++    env_config: null
++
++  blackjack:
++    env_type: blackjack
++    max_actions_per_traj: 10
++    env_instruction: |
++      You are playing Blackjack against a dealer. The dealer must hit on 16 or less and stand on 17 or more.
++      Choose either Stick or Hit. Respond with a single action.
++      Example: <answer>Hit</answer>
++    max_tokens: 64
++    env_config: null
++
++  rubikscube:
++    env_type: rubikscube
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a 2x2 Rubik's Cube (Pocket Cube). The goal is to restore the cube so that each of the faces consists of a single, unique color.
++      Available actions use standard Singmaster notation for face rotations: U, U', D, D', L, L', R, R', F, F', B, B'.
++      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
++      - Modifiers: A letter alone means 90° clockwise (e.g., 'R'). A letter with prime (') means 90° counter-clockwise (e.g., "R'").
++      Respond with a sequence of actions separated by "||".
++      Example: <answer>U</answer>
++    max_tokens: 96
++    env_config:
++      scramble_depth: 3
++      max_steps: 20
++      render_mode: "text"
++  
++  SimpleSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 20
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 2x2 box contains the numbers 1-4 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format, for example: <answer>place 1 at row 2 col 3</answer> or <answer>1,2,3</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 4
++      difficulty: "easy"
++      render_format: "with_feedback"
++      show_conflicts: false
++      show_valid_numbers: false
++      max_steps: 20
++
++  MediumSudoku:
++    env_type: sudoku
++    max_actions_per_traj: 30
++    env_instruction: |
++      You are solving a Sudoku puzzle. Fill in the grid so that every row, column, and 3x3 box contains the numbers 1-9 without repetition.
++      Initial cells are shown in [brackets] and cannot be modified. Empty cells are shown as dots (.).
++      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer> or <answer>2,3,5</answer>
++      The environment will provide feedback on valid/invalid moves and show conflicts if any occur.
++    max_tokens: 150
++    parallel_friendly: false
++    max_workers: 32
++    env_config:
++      grid_size: 9
++      difficulty: "medium"
++      render_format: "with_feedback"
++      show_conflicts: true
++      show_valid_numbers: true
++      max_steps: 81
++
+diff --git a/config/eval.yaml b/config/eval.yaml
+index e65f3a5..0802a0d 100644
+--- a/config/eval.yaml
++++ b/config/eval.yaml
+@@ -1,14 +1,15 @@
+ defaults:
+   - envs
+-
++enable_response_mask: True
+ system:
+-  CUDA_VISIBLE_DEVICES: "0"
++  CUDA_VISIBLE_DEVICES: "0,1,2,3"
+ 
+ seed:
+   train: 10000
+   val: 123
+ 
+-model_path: Qwen/Qwen2.5-3B-Instruct
++model_path: /mnt/general/wanghy/RAGEN/saves/qwen3b_it_fromit_think_sudoku_sequence_multitask/global_step_200/qwen2.5_3b_actor_hf
++# /mnt/general/share/model/tyzhu/SPA-frozenlake-qwen2.5-1.5b-instruct
+ 
+ lora:
+   rank: 0
+@@ -24,17 +25,18 @@ actor_rollout_ref:
+   rollout:
+     name: vllm
+     log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+-    tensor_model_parallel_size: 1
++    tensor_model_parallel_size: 4
+     dtype: bfloat16
+-    max_model_len: 3600
++    max_model_len: 7200
+     prompt_length: 1
+     response_length: 400
+-    gpu_memory_utilization: 0.9
++    gpu_memory_utilization: 0.7
+     max_num_batched_tokens: 8192
+     enforce_eager: True
+     free_cache_engine: True
+     enable_chunked_prefill: False
+     disable_log_stats: False
++    do_sample: True
+     val_kwargs:
+       do_sample: True
+       temperature: 0.5
+@@ -44,9 +46,9 @@ actor_rollout_ref:
+ 
+ agent_proxy:
+   max_context_window: -1
+-  max_turn: 5
++  max_turn: 25
+   action_sep: "||"
+-  max_actions_per_turn: 2
++  max_actions_per_turn: 1
+   use_turn_scores: False
+   enable_think: True
+   reward_normalization:
+@@ -59,14 +61,14 @@ es_manager:
+     env_groups: 8
+     group_size: 16
+     env_configs:
+-      tags: ["CoordSokoban"]
++      tags: ["rubikscube"]
+       n_groups: [8]
+   val:
+-    env_groups: 32
+-    group_size: 16
++    env_groups: 256
++    group_size: 1
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [32]
++      tags: ["rubikscube"]
++      n_groups: [256]
+ 
+ ctx_manager:
+   generation:
+diff --git a/config/evaluate_api_llm.yaml b/config/evaluate_api_llm.yaml
+index b21c95e..bcbf206 100644
+--- a/config/evaluate_api_llm.yaml
++++ b/config/evaluate_api_llm.yaml
+@@ -1,8 +1,11 @@
++# export OPENAI_BASE_URL="https://api.ohmygpt.com/v1"
++# export OPENAI_API_KEY="sk-o4sMxBkN5BB100C4D4a3T3BlBkFJF7791CA39EA14ca98041"
++# python -m ragen.eval_api hydra.searchpath='[file://./verl/verl/trainer/config]'
+ defaults:
+   - base # this is a symbolic link to the verl/verl/trainer/config/ppo_trainer.yaml file
+ 
+ model_config:
+-  model_name: gpt-4o # should be registered in model_info
++  model_name: TA/openai/gpt-oss-120b # should be registered in model_info
+   max_concurrency: 16
+ 
+ model_info:
+@@ -24,9 +27,9 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+-  gpt-4o:
++  gpt-4o-mini:
+     provider_name: openai
+-    model_name: gpt-4o
++    model_name: gpt-4o-mini
+     generation_kwargs:
+       temperature: 0
+       max_tokens: 512 # max_completion_tokens if o1-mini
+@@ -36,21 +39,42 @@ model_info:
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  ark-deepseek-v3-250324:
++    provider_name: openai
++    model_name: ark-deepseek-v3-250324
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
+   deepseek-v3:
+     provider_name: deepseek
+     model_name: deepseek-chat
+     generation_kwargs:
+       temperature: 0
+       max_completion_tokens: 512
++  glm-4.6:
++    provider_name: openai
++    model_name: glm-4.6
++    generation_kwargs:
++      temperature: 0
++      max_completion_tokens: 512
++  TA/openai/gpt-oss-120b:
++    provider_name: openai
++    model_name: TA/openai/gpt-oss-120b
++    generation_kwargs:
++      temperature: 0
++      max_tokens: 8192
++      # max_retries: 5
+ 
+-
+-
++agent_proxy:
++  max_turn: 5
+ es_manager:
+   val:
+-    env_groups: 256
++    env_groups: 32
+     group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
+     env_configs:
+-      tags: ["CoordSokoban"]
+-      n_groups: [256] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
++      tags: ["rubikscube"]
++      n_groups: [32] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
+ 
+ 
++rollout:
++  max_model_len: 7200
+
+Submodule external/webshop-minimal contains modified content
+diff --git a/external/webshop-minimal/requirements.txt b/external/webshop-minimal/requirements.txt
+index 5a1b04f..238ed5a 100644
+--- a/external/webshop-minimal/requirements.txt
++++ b/external/webshop-minimal/requirements.txt
+@@ -4,7 +4,7 @@ flask
+ html2text
+ rank_bm25
+ pyserini
+-faiss-cpu
++faiss-gpu
+ thefuzz
+ gdown
+ spacy
+diff --git a/ragen/env/__init__.py b/ragen/env/__init__.py
+index b0f3461..bc6838e 100644
+--- a/ragen/env/__init__.py
++++ b/ragen/env/__init__.py
+@@ -12,6 +12,14 @@ from .metamathqa.env import MetaMathQAEnv
+ from .metamathqa.config import MetaMathQAEnvConfig
+ from .lean.config import LeanEnvConfig
+ from .lean.env import LeanEnv
++from .game_2048.config import Game2048EnvConfig
++from .game_2048.env import Game2048Env
++from .blackjack.config import BlackjackEnvConfig
++from .blackjack.env import BlackjackEnv
++from .rubikscube.config import RubiksCube2x2Config
++from .rubikscube.env import RubiksCube2x2Env
++from .sudoku.config import SudokuEnvConfig
++from .sudoku.env import SudokuEnv
+ 
+ 
+ REGISTERED_ENVS = {
+@@ -22,6 +30,10 @@ REGISTERED_ENVS = {
+     # 'alfworld': AlfredTXTEnv,
+     'metamathqa': MetaMathQAEnv,
+     'lean': LeanEnv,
++    'game_2048': Game2048Env,
++    'blackjack': BlackjackEnv,
++    'rubikscube': RubiksCube2x2Env,
++    'sudoku': SudokuEnv,
+ }
+ 
+ REGISTERED_ENV_CONFIGS = {
+@@ -32,6 +44,10 @@ REGISTERED_ENV_CONFIGS = {
+     # 'alfworld': AlfredEnvConfig,
+     'metamathqa': MetaMathQAEnvConfig,
+     'lean': LeanEnvConfig,
++    'game_2048': Game2048EnvConfig,
++    'blackjack': BlackjackEnvConfig,
++    'rubikscube': RubiksCube2x2Config,
++    'sudoku': SudokuEnvConfig,
+ }
+ 
+ try:
+diff --git a/ragen/env/bandit/env.py b/ragen/env/bandit/env.py
+index c59f142..53d0f62 100644
+--- a/ragen/env/bandit/env.py
++++ b/ragen/env/bandit/env.py
+@@ -11,6 +11,37 @@ Game Rules:
+ 4. Based on the symbolic meaning of their names, which arm do you think is more likely to give higher rewards on average? Choose between {name_a} and {name_b}, and output like <answer> {name_a} </answer> or <answer> {name_b} </answer>.
+ """
+ 
++# INIT_PROMPT = """You are playing a bandit game. Goal: Maximize your total reward by choosing which arm to pull.
++# Game Rules:
++# 1. There are 2 arms, named {name_a} and {name_b}.
++# 2. Each arm has its own reward distribution, related to their names.
++# 3. You need to analyze the symbolic meaning of each arm's name to guess their reward potential.
++
++# Example answer format:
++# <think>
++# <observation>
++# Arm A is named \"{name_a}\". Symbolically, this implies [analysis of name A].
++# Arm B is named \"{name_b}\". Symbolically, this implies [analysis of name B].
++# </observation>
++# Based on the analysis, {name_a} seems to represent risk/low value, while {name_b} implies wealth/stability.
++# <prediction>
++# If I pull {name_b}, I expect a higher average reward because [reasoning].
++# </prediction>
++# </think>
++# <answer> {name_b} </answer>
++
++# A sample full output is as follows:
++# <think>
++# <observation>
++# Arm A is named "Rotten Apple". Symbolically, this implies decay and zero value.
++# Arm B is named "Golden Chalice". Symbolically, this implies treasure and high value.
++# </observation>
++# Comparing the two, the Golden Chalice is clearly superior in potential value.
++# <prediction>
++# Pulling "Golden Chalice" will likely yield a high positive reward, whereas "Rotten Apple" might give zero or negative reward.
++# </prediction>
++# </think>
++# <answer> Golden Chalice </answer>"""
+ 
+ class BanditEnv(BaseDiscreteActionEnv, gym.Env):
+     class_counter = 0
+diff --git a/ragen/env/frozen_lake/config.py b/ragen/env/frozen_lake/config.py
+index de054f4..9950c34 100644
+--- a/ragen/env/frozen_lake/config.py
++++ b/ragen/env/frozen_lake/config.py
+@@ -8,7 +8,7 @@ class FrozenLakeEnvConfig:
+     size: int = 4
+     p: float = 0.9
+     success_rate: float = 0.8
+-    is_slippery: bool = True
++    is_slippery: bool = False
+     map_seed: Optional[int] = None
+     render_mode: str = "text"
+     observation_format: str = "grid"
+diff --git a/ragen/env/frozen_lake/env.py b/ragen/env/frozen_lake/env.py
+index 9ee3add..eef11d8 100644
+--- a/ragen/env/frozen_lake/env.py
++++ b/ragen/env/frozen_lake/env.py
+@@ -13,6 +13,7 @@ from ragen.env.base import BaseDiscreteActionEnv
+ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+     def __init__(self, config: FrozenLakeEnvConfig = FrozenLakeEnvConfig()):
+         # Using mappings directly from config
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.GRID_LOOKUP = config.grid_lookup
+         self.ACTION_LOOKUP = config.action_lookup
+@@ -95,6 +96,7 @@ class FrozenLakeEnv(BaseDiscreteActionEnv, GymFrozenLakeEnv):
+ if __name__ == "__main__":
+     import matplotlib.pyplot as plt
+     config = FrozenLakeEnvConfig(size=4, is_slippery=True)
++    # import pdb;pdb.set_trace()
+     env = FrozenLakeEnv(config)
+     print(env.reset())
+     while True:
+diff --git a/ragen/env/sokoban/env.py b/ragen/env/sokoban/env.py
+index 17cd636..6922b54 100644
+--- a/ragen/env/sokoban/env.py
++++ b/ragen/env/sokoban/env.py
+@@ -33,6 +33,7 @@ class SokobanEnv(BaseDiscreteActionEnv, GymSokobanEnv):
+     def reset(self, seed=None, mode=None):
+         try:
+             with all_seed(seed):
++                # import pdb;pdb.set_trace()
+                 self.room_fixed, self.room_state, self.box_mapping, action_sequence = generate_room(
+                     dim=self.dim_room,
+                     num_steps=self.num_gen_steps,
+diff --git a/ragen/llm_agent/agent_proxy.py b/ragen/llm_agent/agent_proxy.py
+index c25b50d..f6c4088 100644
+--- a/ragen/llm_agent/agent_proxy.py
++++ b/ragen/llm_agent/agent_proxy.py
+@@ -287,6 +287,7 @@ def main(config):
+ 	proxy = LLMAgentProxy(config, actor_wg, tokenizer)
+ 	import time
+ 	start_time = time.time()
++	# import pdb;pdb.set_trace()
+ 	rollouts = proxy.rollout(
+ 		DataProto(
+ 			batch=None,
+diff --git a/ragen/llm_agent/base_llm.py b/ragen/llm_agent/base_llm.py
+index 358a2eb..b6f7797 100644
+--- a/ragen/llm_agent/base_llm.py
++++ b/ragen/llm_agent/base_llm.py
+@@ -33,6 +33,7 @@ class OpenAIProvider(LLMProvider):
+             raise ValueError("OpenAI API key not provided and not found in environment variables")
+         
+         self.client = AsyncOpenAI(api_key=self.api_key)
++        # import pdb;pdb.set_trace()
+     
+     async def generate(self, messages: List[Dict[str, str]], **kwargs) -> LLMResponse:
+         if "o1-mini" in self.model_name:
+@@ -46,10 +47,7 @@ class OpenAIProvider(LLMProvider):
+         )
+         if response.choices[0].finish_reason in ['length', 'content_filter']:
+             raise ValueError("Content filtered or length exceeded")
+-        return LLMResponse(
+-            content=response.choices[0].message.content,
+-            model_name=response.model
+-        )
++        return LLMResponse(content=response.choices[0].message.content,model_name=response.model)
+ 
+ class DeepSeekProvider(LLMProvider):
+     """DeepSeek API provider implementation"""
+@@ -199,7 +197,7 @@ class ConcurrentLLM:
+         
+         # Queue to store unfinished or failed tasks
+         current_batch = messages_list.copy()
+-        max_retries = kwargs.get("max_retries", 100)
++        max_retries = kwargs.get("max_retries", 10)
+         retry_count = 0
+         
+         while current_batch and retry_count < max_retries:
+diff --git a/ragen/llm_agent/ctx_manager.py b/ragen/llm_agent/ctx_manager.py
+index 905247a..794bd2a 100644
+--- a/ragen/llm_agent/ctx_manager.py
++++ b/ragen/llm_agent/ctx_manager.py
+@@ -88,6 +88,7 @@ class ContextManager:
+         Initialize the ContextManager.
+         Processor is used to process the image data.
+         """
++        # import pdb;pdb.set_trace()
+         self.config = config
+         self.tokenizer = tokenizer
+         self.processor = processor
+@@ -311,7 +312,7 @@ class ContextManager:
+ 
+             llm_input_texts.append(text_with_prompt)
+             messages_list.append(messages)
+-
++        # import pdb;pdb.set_trace()
+         inputs = self.tokenizer(llm_input_texts, return_tensors="pt", padding=True, padding_side="left", truncation=False) # We have truncated previously, truncation in tokenizer may cause issues.
+         input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
+         position_ids = (attention_mask.cumsum(dim=-1) - 1).clamp(min=0)
+@@ -331,7 +332,6 @@ class ContextManager:
+             "position_ids": position_ids,
+             "responses": input_ids[:, 1:], # remove the first token
+         }, batch_size=input_ids.shape[0])
+-
+         if prepare_for_update:
+             llm_inputs.batch["loss_mask"] = loss_mask # remove the first token
+             llm_inputs.batch["rm_scores"] = normalized_score_tensor # remove the first token
+@@ -353,6 +353,30 @@ class ContextManager:
+                 key: np.sum(value) / self.env_nums[key.split("/")[0]]
+                 for key, value in metrics.items()
+             }
++            # Derived metrics for wandb logging
++            try:
++                # charts/avg_episode_return: average across all 2048 env groups
++                ep_keys = [k for k in metrics.keys() if k.endswith('/episodic_return')]
++                if len(ep_keys) > 0:
++                    ep_vals = []
++                    for k in ep_keys:
++                        tag = k.split('/')[0]
++                        denom = self.env_nums.get(tag, max(1, len(metrics[k])))
++                        ep_vals.append(float(np.sum(metrics[k]) / denom))
++                    mean_metrics["charts/avg_episode_return"] = float(np.mean(ep_vals))
++            except Exception:
++                pass
++            try:
++                # rollout/max_tile: max over all envs in this batch
++                tile_keys = [k for k in metrics.keys() if k.endswith('/max_tile')]
++                if len(tile_keys) > 0:
++                    tile_vals = []
++                    for k in tile_keys:
++                        tile_vals.extend(metrics[k])
++                    if len(tile_vals) > 0:
++                        mean_metrics["rollout/max_tile"] = int(np.max(tile_vals))
++            except Exception:
++                pass
+             for key, values in metrics.items():
+                 if not isinstance(values, list):
+                     continue
+@@ -378,6 +402,7 @@ class ContextManager:
+             
+         env_ids = lm_outputs.non_tensor_batch['env_ids']
+         env_inputs = []
++        # import pdb;pdb.set_trace()
+         for env_id, response in zip(env_ids, responses):
+             llm_response, actions = self._parse_response(response)
+             env_inputs.append({
+diff --git a/ragen/llm_agent/es_manager.py b/ragen/llm_agent/es_manager.py
+index b87a2b3..1a49d3e 100644
+--- a/ragen/llm_agent/es_manager.py
++++ b/ragen/llm_agent/es_manager.py
+@@ -128,18 +128,40 @@ class EnvStateManager:
+         env_outputs: List[Dict]
+             {env_id: int, history: List[Dict][{state: str, actions: List[str], reward: float, info: Dict, llm_response: str, llm_raw_response: str, (Optional)images: List[PIL.Image.Image]}]}
+         """
++        # def _execute_actions(env, actions):
++        #     acc_reward, turn_info, turn_done = 0, {}, False
++        #     executed_actions = []
++        #     for action in actions:
++        #         _, reward, done, info = env.step(action)
++        #         acc_reward += reward
++        #         turn_info.update(info) # NOTE: currently use last info for multi-action
++        #         executed_actions.append(action)
++        #         if done:
++        #             turn_done = True
++        #             break
++            
++        #     return acc_reward, turn_info, turn_done, executed_actions
+         def _execute_actions(env, actions):
+-            acc_reward, turn_info, turn_done = 0, {}, False
++            acc_reward, turn_info, turn_done = 0.0, {}, False
++            raw_acc_reward = 0.0
+             executed_actions = []
+             for action in actions:
+                 _, reward, done, info = env.step(action)
+-                acc_reward += reward
++                acc_reward += float(reward)
++                try:
++                    raw_acc_reward += float(info.get('raw_reward', 0.0))
++                except Exception:
++                    pass
+                 turn_info.update(info) # NOTE: currently use last info for multi-action
+                 executed_actions.append(action)
+                 if done:
+                     turn_done = True
+                     break
+-            
++            # Overwrite per-turn raw_reward to reflect the sum across all executed actions in this turn
++            try:
++                turn_info['raw_reward'] = float(raw_acc_reward)
++            except Exception:
++                pass
+             return acc_reward, turn_info, turn_done, executed_actions
+ 
+         def _log_env_state(status, history, cur_obs, max_actions_per_traj, executed_actions, all_actions, acc_reward, turn_done, turn_info, env_input):
+@@ -198,6 +220,18 @@ class EnvStateManager:
+                 'success': float(status.terminated and (not status.truncated)),
+                 'num_actions': status.num_actions,
+             }
++            # Add episodic-level metrics
++            # try:
++            #     # Sum of per-turn rewards equals the episodic return (env-shaped reward)
++            #     env_metric['episodic_return'] = float(sum(status.rewards))
++            # except Exception:
++            #     pass
++            try:
++                # Final max tile on the board at the end of the rollout
++                import numpy as _np
++                env_metric['max_tile'] = int(_np.max(entry['env'].grid))
++            except Exception:
++                pass
+             custom_metric = {}
+             for turn in cache['history']:
+                 for k, v in turn.get('info', {}).items():
+@@ -212,6 +246,12 @@ class EnvStateManager:
+                             "Skipping non-numeric metric '%s' with value %r for env %s.",
+                             k, v, entry['tag']
+                         )
++            # Add episodic_return as the SUM of raw_reward across steps (align with CleanRL)
++            try:
++                if 'raw_reward' in custom_metric:
++                    env_metric['episodic_return'] = float(np.sum(custom_metric['raw_reward']))
++            except Exception:
++                pass
+             for k, v in custom_metric.items():
+                 # TODO: Move TURN_LVL_METRICS into the environment
+                 if "webshop" not in cache['tag'].lower() or ("webshop" in cache['tag'].lower() and k in TURN_LVL_METRICS):
+@@ -219,7 +259,12 @@ class EnvStateManager:
+                 else:
+                     env_metric['traj_sum/' + k] = np.sum(v)
+ 
+-
++            try:
++                if 'score' in custom_metric and len(custom_metric['score']) > 0:
++                    env_metric['final_score'] = float(custom_metric['score'][-1])
++            except Exception:
++                pass
++                
+             cache['history'][-1]['metrics'] = custom_metric
+             env_metric = {f"{entry['tag']}/{k}": v for k, v in env_metric.items()}
+             cache['metrics'] = env_metric
+diff --git a/requirements.txt b/requirements.txt
+index 2fd756e..9b521bf 100644
+--- a/requirements.txt
++++ b/requirements.txt
+@@ -7,7 +7,6 @@ accelerate
+ codetiming
+ datasets
+ dill
+-flash-attn==2.7.4.post1
+ hydra-core
+ numpy
+ pandas
+@@ -15,19 +14,19 @@ pybind11
+ ray>=2.10
+ tensordict>=0.8.0,<0.9.0
+ transformers
+-vllm==0.8.2
++vllm==0.8.5
+ wandb
+ gymnasium
+ gymnasium[toy-text]
+ 
+ pyarrow>=15.0.0
+ pylatexenc
+-torchdata
++# torchdata
+ debugpy
+ 
+ together
+ anthropic
+-faiss-cpu==1.11.0
++faiss-gpu
+ 
+ # This is optional, but needs to be installed with main requirements if you want to use webshop
+ -r external/webshop-minimal/requirements.txt
+diff --git a/scripts/setup_ragen.sh b/scripts/setup_ragen.sh
+index f9a7cd9..85a93c4 100644
+--- a/scripts/setup_ragen.sh
++++ b/scripts/setup_ragen.sh
+@@ -94,10 +94,10 @@ main() {
+         pip install torch==2.5.0 --index-url https://download.pytorch.org/whl/cu124
+         
+         print_step "Installing flash-attention..."
+-        pip3 install flash-attn==2.7.4.post1 --no-build-isolation
++        # pip3 install flash-attn==2.7.4.post1 --no-build-isolation
+     else
+         print_step "Installing PyTorch without CUDA support..."
+-        pip install torch==2.5.0
++        pip install torch==2.4.0
+     fi
+     
+     # Install remaining requirements
+@@ -137,8 +137,8 @@ main() {
+     conda install conda-forge::gdown
+     mkdir -p external/webshop-minimal/webshop_minimal/data/full
+     cd external/webshop-minimal/webshop_minimal/data/full
+-    gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
+-    gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
++    # gdown https://drive.google.com/uc?id=1A2whVgOO0euk5O13n2iYDM0bQRkkRduB # items_shuffle
++    # gdown https://drive.google.com/uc?id=1s2j6NgHljiZzQNL3veZaAiyW_qDEgBNi # items_ins_v2
+     cd ../../../../..
+ 
+     echo -e "${GREEN}Installation completed successfully!${NC}"
+diff --git a/train.py b/train.py
+index 34af78b..00c2117 100644
+--- a/train.py
++++ b/train.py
+@@ -132,7 +132,7 @@ def add_dependency_and_validate_config(config):
+         f"micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node}) must be less than or equal to ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size})"
+     assert config.actor_rollout_ref.actor.ppo_mini_batch_size % (config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node) == 0, \
+         f"ppo_mini_batch_size ({config.actor_rollout_ref.actor.ppo_mini_batch_size}) must be divisible by micro_batch_size_per_gpu * n_gpus_per_node ({config.micro_batch_size_per_gpu * config.trainer.n_gpus_per_node})"
+-    assert "qwen" in config.model_path.lower() or (not config.enable_response_mask), \
++    assert "qwen" in config.model_path.lower() or "llama-3" in config.model_path.lower() or (not config.enable_response_mask), \
+         "response mask is currently only supported for qwen models"
+     assert len(str(config.system.CUDA_VISIBLE_DEVICES).split(',')) == config.trainer.n_gpus_per_node, \
+         f"CUDA_VISIBLE_DEVICES ({config.system.CUDA_VISIBLE_DEVICES}) must have the same number of GPUs as n_gpus_per_node ({config.trainer.n_gpus_per_node})"
+diff --git a/train_all.sh b/train_all.sh
+index 0157306..33035f5 100755
+--- a/train_all.sh
++++ b/train_all.sh
+@@ -6,246 +6,257 @@ USE_GRPO="algorithm.adv_estimator=grpo"
+ USE_PPO="algorithm.adv_estimator=gae" # by default.
+ USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+ 
++
++# python train.py --config-name _8_2048 system.CUDA_VISIBLE_DEVICES="'0,1,2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=game_2048 $USE_PPO $USE_BASE 
++
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-1 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-1.5B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++# python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'4,5'" trainer.project_name=ragen_latest_qwen_25_3b_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-3B-Instruct  trainer.experiment_name=rubikscube-2 $USE_PPO $USE_BASE 
++python train.py --config-name _10_rubikscube system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_05B_it trainer.n_gpus_per_node=2 model_path=Qwen/Qwen2.5-0.5B-Instruct  trainer.experiment_name=rubikscube-3 $USE_PPO $USE_BASE 
++
+ # Section 3.1&3.2 - General Observations
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozen_lake-ppo $USE_PPO $USE_BASE &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozen_lake-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'6,7'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=bandit-ppo-multitask $USE_PPO $USE_BASE &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="'7'" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-ppo-frommlp $USE_PPO $USE_BASE 
+ 
+-# Section 4.1 - Filtering and critic
+-# 0.25
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+-
+-wait
+-
+-# 0.5
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
+-
+-# 0.75
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+-
+-wait
+-
+-# Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="'0,1'" trainer.project_name=ragen_latest_qwen_25_15b_it  trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=sokoban-ppo-box1-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-grpo $USE_GRPO $USE_BASE &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'2,3'" trainer.project_name=ragen_latest_qwen_25_15b_it trainer.n_gpus_per_node=2 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-wait
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.project_name=ragen_latest_qwen_25_3b_it  trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_multitask  trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="'4,5,6,7'" trainer.n_gpus_per_node=4 model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask trainer.experiment_name=frozen_lake-ppo-slippery-multitask $USE_PPO $USE_BASE 
+ 
+-# Section 5.1 - Reasoning Helps Generalization
++# # Section 4.1 - Filtering and critic
++# # 0.25
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.25 actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.25 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
+-    agent_proxy.enable_think=False
++# # 0.5
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-ppo-rolloutfilter0.5 actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.5 $USE_GRPO &
++
++# # 0.75
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-ppo-rolloutfilter0.75 actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozen_lake-ppo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozen_lake-grpo actor_rollout_ref.rollout.rollout_filter_ratio=0.75 $USE_GRPO &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader"
++# wait
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
+-    custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
+-    custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
+-    custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
+-    custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
+-    agent_proxy.enable_think=False
++# # Section 4.2 - Ablation on Critic/ClipHigh/KL. Start from Basic and add more components. The best setting for StarPO in agent is rollout_filter+Critic+Cliphigh+NoKL
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=frozenlake-base-grpo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_GRPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=frozenlake-base-ppo algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=frozenlake-base-ppo-cliphigh algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.28 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=frozenlake-base-ppo-nokl algorithm.kl_ctrl.kl_coef=0.000 actor_rollout_ref.actor.kl_loss_coef=0.000 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1 $USE_PPO &
+ 
++# wait
+ 
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
++# # Section 5.1 - Reasoning Helps Generalization
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=bandit-generalization \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian"
+ 
+-# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=bandit-generalization-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Engineer" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Teacher" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Trader" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Librarian" \
++#     agent_proxy.enable_think=False
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=bandit-generalization-rev \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader"
+ 
+-# COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=bandit-generalization-rev-nothink \
++#     custom_envs.Bandit.env_config.lo_arm_name="Teacher" \
++#     custom_envs.Bandit.env_config.hi_arm_name="Engineer" \
++#     custom_envs.BanditTest.env_config.lo_arm_name="Librarian" \
++#     custom_envs.BanditTest.env_config.hi_arm_name="Trader" \
++#     agent_proxy.enable_think=False
+ 
+-wait
+ 
+-# Section 5.2 - what leads to better reasoning?
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization micro_batch_size_per_gpu=8 model_path=Qwen/Qwen2.5-1.5B-Instruct&
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-generalization-nothink $SOKOBAN_GENERALIZATION_CONFIG agent_proxy.enable_think=False &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
++# # SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=128 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SokobanDifferentGridVocab] es_manager.val.env_configs.n_groups=[128]"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization $SOKOBAN_GENERALIZATION_CONFIG &
+ 
+-wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
++# # COMPOSITIONALITY_CONFIG="es_manager.train.env_groups=16 es_manager.train.env_configs.tags=[Bandit,SimpleSokoban] es_manager.train.env_configs.n_groups=[8,8] es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[Bandit,SimpleSokoban,LargerSokoban,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128] actor_rollout_ref.rollout.rollout_filter_ratio=1" # NOTE that we don't filter out low-var rollout in this setting
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=compositional-generalization $COMPOSITIONALITY_CONFIG &
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=compositional-generalization-nothink $COMPOSITIONALITY_CONFIG agent_proxy.enable_think=False &
+ 
++# wait
+ 
+-wait
++# # Section 5.2 - what leads to better reasoning?
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B-Instruct &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct trainer.n_gpus_per_node=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-instruct $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B-Instruct trainer.n_gpus_per_node=4 &
+ 
+-# Section 6.1 varying action count
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
++# wait
+ 
+-# section 6.2 Varying prompt diversity
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-generalization-qwen2.5-0.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-0.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-1.5B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-3b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5,6,7\" trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-7B &
+ 
+-wait 
++# wait
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\"  trainer.n_gpus_per_node=4 trainer.n_gpus_per_node=4 trainer.experiment_name=sokoban-generalization-qwen2.5-7b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\"  trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-generalization-qwen2.5-1.5b-r1 $SOKOBAN_GENERALIZATION_CONFIG model_path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
++# wait
+ 
++# # Section 6.1 varying action count
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-action-count-1 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=5 custom_envs.LargerSokoban.max_actions_per_traj=5 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=5 custom_envs.FrozenLake.max_actions_per_traj=5 agent_proxy.max_actions_per_turn=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-action-count-2 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=10 custom_envs.LargerSokoban.max_actions_per_traj=10 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=10 custom_envs.FrozenLake.max_actions_per_traj=10 agent_proxy.max_actions_per_turn=2 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-action-count-3 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=15 custom_envs.LargerSokoban.max_actions_per_traj=15 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=15 custom_envs.FrozenLake.max_actions_per_traj=15 agent_proxy.max_actions_per_turn=3 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-action-count-4 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=20 custom_envs.LargerSokoban.max_actions_per_traj=20 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=20 custom_envs.FrozenLake.max_actions_per_traj=20 agent_proxy.max_actions_per_turn=4 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-action-count-5 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=25 custom_envs.LargerSokoban.max_actions_per_traj=25 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=25 custom_envs.FrozenLake.max_actions_per_traj=25 agent_proxy.max_actions_per_turn=5 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-action-count-6 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=30 custom_envs.LargerSokoban.max_actions_per_traj=30 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=30 custom_envs.FrozenLake.max_actions_per_traj=30 agent_proxy.max_actions_per_turn=6 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-action-count-7 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=35 custom_envs.LargerSokoban.max_actions_per_traj=35 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=35 custom_envs.FrozenLake.max_actions_per_traj=35 agent_proxy.max_actions_per_turn=7 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-action-count-8 $SOKOBAN_GENERALIZATION_CONFIG custom_envs.SimpleSokoban.max_actions_per_traj=40 custom_envs.LargerSokoban.max_actions_per_traj=40 custom_envs.SokobanDifferentGridVocab.max_actions_per_traj=40 custom_envs.FrozenLake.max_actions_per_traj=40 agent_proxy.max_actions_per_turn=8 &
+ 
++# # section 6.2 Varying prompt diversity
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-prompt-diversity-4 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=4 es_manager.train.group_size=32 es_manager.train.env_configs.n_groups=[4] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="1" trainer.experiment_name=sokoban-prompt-diversity-8 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="2" trainer.experiment_name=sokoban-prompt-diversity-16 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=8 es_manager.train.env_configs.n_groups=[16] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="3" trainer.experiment_name=sokoban-prompt-diversity-32 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=32 es_manager.train.group_size=4 es_manager.train.env_configs.n_groups=[32] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="4" trainer.experiment_name=sokoban-prompt-diversity-64 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=64 es_manager.train.group_size=2 es_manager.train.env_configs.n_groups=[64] &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="5" trainer.experiment_name=sokoban-prompt-diversity-128 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=128 es_manager.train.group_size=1 es_manager.train.env_configs.n_groups=[128] &
+ 
++# wait 
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="6" trainer.experiment_name=sokoban-online-2 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=16 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[16] trainer.total_training_steps=100 trainer.test_freq=5 &
+ 
+-# Extension: Training 7B reasoning model
+-SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="7" trainer.experiment_name=sokoban-online-5 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=40 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[40] trainer.total_training_steps=40 trainer.test_freq=2 &
+ 
+-python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES="0" trainer.experiment_name=sokoban-online-10 $SOKOBAN_GENERALIZATION_CONFIG es_manager.train.env_groups=80 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[80] trainer.total_training_steps=80 trainer.test_freq=1 &
+ 
+-# trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
+-python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
++# # Extension: Training 7B reasoning model
++# SOKOBAN_GENERALIZATION_CONFIG="es_manager.val.env_groups=512 es_manager.val.group_size=1 es_manager.val.env_configs.tags=[SimpleSokoban,LargerSokoban,SokobanDifferentGridVocab,FrozenLake] es_manager.val.env_configs.n_groups=[128,128,128,128]"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 trainer.experiment_name=sokoban-generalization-qwen2.5-3b-instruct-largescale $SOKOBAN_GENERALIZATION_CONFIG model_path=Qwen/Qwen2.5-3B-Instruct es_manager.train.env_groups=8 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[8] micro_batch_size_per_gpu=8 ppo_mini_batch_size=64 actor_rollout_ref.rollout.response_length=1024 actor_rollout_ref.rollout.max_model_len=6400 trainer.test_freq=5 actor_rollout_ref.rollout.max_num_batched_tokens=24000 micro_batch_size_per_gpu=2 actor_rollout_ref.rollout.rollout_filter_ratio=1 &
+ 
++# python -m ragen.llm_agent.agent_proxy  model_path=Qwen/Qwen2.5-3B-Instruct system.CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\" trainer.n_gpus_per_node=8 actor_rollout_ref.rollout.tensor_model_parallel_size=4 actor_rollout_ref.rollout.response_length=2048 actor_rollout_ref.rollout.max_model_len=12800 
+ 
++# # trainer.save_freq=50 trainer.default_local_dir=/mnt/local/cache/exp_name
+ 
+-# USE_PPO="algorithm.adv_estimator=gae" # by default.
+-# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/sokoban-generalization &
+ 
++# python train.py --config-name _1_bandit system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=bandit-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/bandit-generalization &
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
++# python train.py --config-name _3_frozen_lake system.CUDA_VISIBLE_DEVICES=\"1\" trainer.n_gpus_per_node=1 trainer.experiment_name=frozenlake-final enable_response_mask=True trainer.total_training_steps=500 trainer.save_freq=50 trainer.default_local_dir=/mnt/local/ragen_checkpoints/frozenlake-generalization &
+ 
+-# enable_response_mask: False
+-# grpo_advantage_length_weight: True
+ 
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+ 
+-# extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
+-python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
++# # USE_PPO="algorithm.adv_estimator=gae" # by default.
++# # USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# # python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0\" trainer.n_gpus_per_node=1 trainer.experiment_name=sokoban-ppo $USE_PPO $USE_BASE ppo_mini_batch_size=64 enable_response_mask=True &
+ 
+-# extension: bi-level gae
+-python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
+-    system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+-    model_path=Qwen/Qwen2.5-0.5B-Instruct \
+-    algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
+-    agent_proxy.use_turn_scores=True \
+-    actor_rollout_ref.rollout.tp_size_check=False
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std &
+ 
+-# extension: webshop
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # enable_response_mask: False
++# # grpo_advantage_length_weight: True
+ 
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 trainer.experiment_name=sokoban-s-grpo-1-5b algorithm.adv_estimator=grpo agent_proxy.reward_normalization.method=mean_std agent_proxy.max_actions_per_turn=5 custom_envs.SimpleSokoban.max_actions_per_traj=25 enable_response_mask=True grpo_advantage_length_weight=False model_path=Qwen/Qwen2.5-1.5B-Instruct &
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: 7B with lora. Currently NOT recommended to use lora: within current version of vllm, this could result in rollouts slower than non-lora by 100%
++# python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 model_path=Qwen/Qwen2.5-7B-Instruct trainer.experiment_name=sokoban_7b_instruct_lora_newversion lora.rank=16  
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # extension: bi-level gae
++# python train.py trainer.experiment_name=sokoban-bi-level-gae-final \
++#     system.CUDA_VISIBLE_DEVICES=\"0,1,2,3\" trainer.n_gpus_per_node=4 actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
++#     model_path=Qwen/Qwen2.5-0.5B-Instruct \
++#     algorithm.bi_level_gae=True algorithm.high_level_gamma=0.95 \
++#     agent_proxy.use_turn_scores=True \
++#     actor_rollout_ref.rollout.tp_size_check=False
+ 
+ 
+-# normal:sokoban
+-# extension: sokoban
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
+-    trainer.nnodes=1 &
++# # extension: webshop
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
+ 
+ 
+-USE_GRPO="algorithm.adv_estimator=grpo" # by default.
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
+-    trainer.nnodes=1 &
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
+ 
+-# StarPO ppo
+-USE_PPO="algorithm.adv_estimator=gae" # by default.
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+-# StarPO grpo
+-USE_GRPO="algorithm.adv_estimator=grpo"
+-USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
+-MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
+-    trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
+-    es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
+-    trainer.nnodes=1 &
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _6_webshop system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=webshop-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++
++# # normal:sokoban
++# # extension: sokoban
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo-s $USE_PPO \
++#     trainer.nnodes=1 &
++
++
++# USE_GRPO="algorithm.adv_estimator=grpo" # by default.
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"2,3\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo-s $USE_GRPO \
++#     trainer.nnodes=1 &
++
++# # StarPO ppo
++# USE_PPO="algorithm.adv_estimator=gae" # by default.
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"4,5\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-ppo $USE_PPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
++
++# # StarPO grpo
++# USE_GRPO="algorithm.adv_estimator=grpo"
++# USE_BASE="algorithm.kl_ctrl.kl_coef=0.001 actor_rollout_ref.actor.kl_loss_coef=0.001 actor_rollout_ref.actor.clip_ratio_high=0.2 actor_rollout_ref.rollout.rollout_filter_ratio=1"
++# MKL_SERVICE_FORCE_INTEL=1 python train.py --config-name _2_sokoban system.CUDA_VISIBLE_DEVICES=\"6,7\" trainer.n_gpus_per_node=2 \
++#     trainer.experiment_name=sokoban-3b-grpo $USE_GRPO $USE_BASE \
++#     es_manager.train.env_groups=2 es_manager.train.group_size=16 es_manager.train.env_configs.n_groups=[2] \
++#     trainer.nnodes=1 &
+ 
+ 
+ 
+-python train.py \
+-    trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+
++# python train.py \
++#     trainer.experiment_name=3b-full-ppo-test system.CUDA_VISIBLE_DEVICES=\"0,1\" trainer.n_gpus_per_node=2
+

wandb/run-20260112_225324-locrg3kr/files/requirements.txt

@@ -0,0 +1,313 @@
+colorama==0.4.6
+psutil==7.1.3
+ragen==0.1
+Farama-Notifications==0.0.4
+MarkupSafe==2.1.5
+PyJWT==2.10.1
+PySocks==1.7.1
+absl-py==2.3.1
+accelerate==1.12.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.2
+aiohttp-cors==0.8.1
+aiosignal==1.4.0
+airportsdata==20250909
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anthropic==0.75.0
+antlr4-python3-runtime==4.9.3
+anyio==4.12.0
+astor==0.8.1
+asttokens==3.0.1
+async-timeout==5.0.1
+attrs==25.4.0
+beautifulsoup4==4.14.3
+black==25.11.0
+blake3==1.0.8
+blinker==1.9.0
+blis==1.3.3
+cachetools==6.2.2
+catalogue==2.0.10
+certifi==2025.11.12
+cffi==2.0.0
+charset-normalizer==3.4.4
+cleantext==1.1.4
+click==8.2.1
+cloudpathlib==0.23.0
+cloudpickle==3.1.2
+codetiming==1.4.0
+colorama==0.4.6
+coloredlogs==15.0.1
+colorful==0.5.8
+compressed-tensors==0.9.3
+confection==0.1.5
+contourpy==1.3.2
+cryptography==46.0.3
+cupy-cuda12x==13.6.0
+cycler==0.12.1
+cymem==2.0.13
+Cython==3.2.2
+datasets==4.4.1
+debugpy==1.8.17
+decorator==5.2.1
+Deprecated==1.3.1
+depyf==0.18.0
+dill==0.4.0
+diskcache==5.6.3
+distlib==0.4.0
+distro==1.9.0
+dnspython==2.8.0
+docstring_parser==0.17.0
+einops==0.8.1
+email-validator==2.3.0
+eval_type_backport==0.2.2
+exceptiongroup==1.3.1
+executing==2.2.1
+faiss-gpu==1.7.2
+fastapi==0.123.1
+fastapi-cli==0.0.16
+fastapi-cloud-cli==0.5.2
+fastar==0.8.0
+fastrlock==0.8.3
+filelock==3.19.1
+flash_attn==2.8.3
+Flask==3.1.2
+flatbuffers==25.9.23
+fonttools==4.61.0
+frozenlist==1.8.0
+fsspec==2025.9.0
+gdown==5.2.0
+gguf==0.17.1
+gitdb==4.0.12
+GitPython==3.1.45
+google-api-core==2.28.1
+google-auth==2.43.0
+googleapis-common-protos==1.72.0
+grpcio==1.76.0
+gym==0.26.2
+gym-notices==0.1.0
+gym-sokoban==0.0.6
+gymnasium==1.2.2
+h11==0.16.0
+hf-xet==1.2.0
+html2text==2025.4.15
+httpcore==1.0.9
+httptools==0.7.1
+httpx==0.28.1
+httpx-sse==0.4.3
+huggingface-hub==0.36.0
+humanfriendly==10.0
+hydra-core==1.3.2
+idna==3.11
+ImageIO==2.37.2
+importlib_metadata==8.0.0
+interegular==0.3.3
+ipython==8.37.0
+itsdangerous==2.2.0
+jedi==0.19.2
+Jinja2==3.1.6
+jiter==0.12.0
+joblib==1.5.2
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+kimina-client==0.2.1
+kiwisolver==1.4.9
+lark==1.2.2
+llguidance==0.7.30
+llvmlite==0.44.0
+lm-format-enforcer==0.10.12
+loguru==0.7.3
+Markdown==3.10
+markdown-it-py==4.0.0
+matplotlib==3.10.7
+matplotlib-inline==0.2.1
+mcp==1.22.0
+mdurl==0.1.2
+mistral_common==1.8.6
+mpmath==1.3.0
+msgpack==1.1.2
+msgspec==0.20.0
+multidict==6.7.0
+multiprocess==0.70.18
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+nest-asyncio==1.6.0
+networkx==3.3
+ninja==1.13.0
+nltk==3.9.2
+numba==0.61.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-cusparselt-cu12==0.6.2
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+omegaconf==2.3.0
+onnxruntime==1.23.2
+openai==2.8.1
+opencensus==0.11.4
+opencensus-context==0.1.3
+opencv-python-headless==4.12.0.88
+opentelemetry-api==1.38.0
+opentelemetry-exporter-otlp==1.26.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+opentelemetry-exporter-prometheus==0.59b0
+opentelemetry-proto==1.26.0
+opentelemetry-sdk==1.38.0
+opentelemetry-semantic-conventions==0.59b0
+opentelemetry-semantic-conventions-ai==0.4.13
+orjson==3.11.4
+outlines==0.1.11
+outlines_core==0.1.26
+packaging==25.0
+pandas==2.3.3
+parso==0.8.5
+partial-json-parser==0.2.1.1.post7
+pathspec==0.12.1
+peft==0.18.0
+pexpect==4.9.0
+pillow==11.3.0
+pip==25.3
+platformdirs==4.5.0
+preshed==3.0.12
+prometheus_client==0.23.1
+prometheus-fastapi-instrumentator==7.1.0
+prompt_toolkit==3.0.52
+propcache==0.4.1
+proto-plus==1.26.1
+protobuf==4.25.8
+psutil==7.1.3
+ptyprocess==0.7.0
+pure_eval==0.2.3
+py-cpuinfo==9.0.0
+py-spy==0.4.1
+pyarrow==22.0.0
+pyasn1==0.6.1
+pyasn1_modules==0.4.2
+pybind11==3.0.1
+pycountry==24.6.1
+pycparser==2.23
+pydantic==2.12.5
+pydantic_core==2.41.5
+pydantic-extra-types==2.10.6
+pydantic-settings==2.12.0
+pygame==2.6.1
+Pygments==2.19.2
+pyjnius==1.7.0
+pylatexenc==2.10
+pyparsing==3.2.5
+pyserini==1.2.0
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.1
+python-json-logger==4.0.0
+python-multipart==0.0.20
+pytokens==0.3.0
+pytz==2025.2
+PyYAML==6.0.3
+pyzmq==27.1.0
+ragen==0.1
+rank-bm25==0.2.2
+RapidFuzz==3.14.3
+ray==2.52.1
+referencing==0.37.0
+regex==2025.11.3
+requests==2.32.5
+rich==14.2.0
+rich-toolkit==0.17.0
+rignore==0.7.6
+rpds-py==0.30.0
+rsa==4.9.1
+safetensors==0.7.0
+scikit-learn==1.7.2
+scipy==1.15.3
+sentencepiece==0.2.1
+sentry-sdk==2.46.0
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.17.0
+smart_open==7.5.0
+smmap==5.0.2
+sniffio==1.3.1
+soupsieve==2.8
+spacy==3.8.11
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.2
+sse-starlette==3.0.3
+stack-data==0.6.3
+starlette==0.50.0
+sympy==1.13.1
+tabulate==0.9.0
+tenacity==9.1.2
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+tensordict==0.8.3
+thefuzz==0.22.1
+thinc==8.3.10
+threadpoolctl==3.6.0
+tiktoken==0.12.0
+together==1.5.31
+tokenizers==0.22.1
+tomli==2.3.0
+torch==2.6.0+cu124
+torchaudio==2.6.0
+torchdata==0.11.0
+torchvision==0.21.0
+tqdm==4.67.1
+traitlets==5.14.3
+transformers==4.57.3
+triton==3.2.0
+typer==0.19.2
+typer-slim==0.20.0
+typing_extensions==4.15.0
+typing-inspection==0.4.2
+tzdata==2025.2
+urllib3==2.5.0
+uvicorn==0.38.0
+uvloop==0.22.1
+verl==0.5.0.dev0
+virtualenv==20.35.4
+vllm==0.8.5
+wandb==0.23.0
+wasabi==1.1.3
+watchfiles==1.1.1
+wcwidth==0.2.14
+weasel==0.4.3
+websockets==15.0.1
+Werkzeug==3.1.4
+wheel==0.45.1
+wrapt==2.0.1
+xformers==0.0.29.post2
+xgrammar==0.1.18
+xxhash==3.6.0
+yarl==1.22.0
+zipp==3.23.0
+typeguard==4.4.4
+tyro==1.0.3
+verl==0.5.0.dev0
+autocommand==2.2.2
+backports.tarfile==1.2.0
+importlib_metadata==8.0.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+jaraco.context==5.3.0
+jaraco.functools==4.0.1
+jaraco.text==3.12.1
+more-itertools==10.3.0
+packaging==24.2
+platformdirs==4.2.2
+tomli==2.0.1
+typeguard==4.3.0
+typing_extensions==4.12.2
+wheel==0.45.1
+zipp==3.19.2

wandb/run-20260112_225324-locrg3kr/files/wandb-metadata.json

@@ -0,0 +1,64 @@
+{
+  "os": "Linux-5.14.0-284.25.1.el9_2.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.19",
+  "startedAt": "2026-01-12T14:53:24.482263Z",
+  "args": [
+    "--track"
+  ],
+  "program": "/mnt/general/wanghy/RAGEN/cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePath": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "codePathLocal": "cleanrl/cleanrl/scout_dqn/dqn_rubikscube.py",
+  "git": {
+    "remote": "https://github.com/mll-lab-nu/RAGEN.git",
+    "commit": "8d73639b99b38265453f898b8d6af7d4af50d56e"
+  },
+  "email": "haoyu-wa22@mails.tsinghua.edu.cn",
+  "root": "/mnt/general/wanghy/RAGEN",
+  "host": "pt-a93c53b7eacd4c2ca962941d3210c0ac-worker-0",
+  "executable": "/root/local/miniconda3/envs/ragen/bin/python",
+  "cpu_count": 64,
+  "cpu_count_logical": 128,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "2576980377600",
+      "used": "3724316672"
+    }
+  },
+  "memory": {
+    "total": "2163639439360"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-cff454b9-f118-535b-5e1f-6d503e78e52e"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9e77c178-188c-f869-780e-858c20a7d322"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c63d047d-a077-6c9a-20aa-d43d7cfb61c9"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6b73ee0d-b923-d26e-4c48-314e4ccbebc7"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "38xx0mqyxzx0knb11tf2de7k74fbhknb"
+}

wandb/run-20260112_225324-locrg3kr/files/wandb-summary.json

@@ -0,0 +1 @@
+{"eval/success_rate":0.99925,"eval/std_return":0.06474532779282224,"_step":400000,"eval/episodes":4000,"train/loss":0.0006442762678489089,"global_step":400000,"_runtime":689.639799334,"eval/avg_return":0.893023,"_timestamp":1.7682302954374332e+09,"train/learning_rate":0.00025,"train/success_rate_100":1,"train/success":1,"perf/SPS":590,"charts/epsilon":0.05,"_wandb":{"runtime":689},"train/episodic_length":1,"train/episodic_return":0.956}