File size: 12,073 Bytes
7d90172 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 | """SPAgent GRPO plugin for deterministic VRBench navigation.
The model-facing interaction contract matches FrozenLake v3: explicit image
placement, one-step ``move`` calls, cumulative image anchors, full conversation
history, and final-answer replay from the initial state.
"""
from __future__ import annotations
import json
import os
import re
import sys
import uuid
from pathlib import Path
from typing import Any, Dict, List
REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from spagent.vrbench_navigation_protocol import ( # noqa: E402
SUPPORTED_TASKS,
clone_state,
create_tool_response,
decode_initial_state,
encode_initial_state,
parse_action_tokens,
render_state,
replay_actions,
step_state,
)
from swift.llm import RolloutInferRequest # noqa: E402
sys.path.insert(0, str(Path(__file__).resolve().parent))
import grpo_filtered_trainer # noqa: E402,F401
from swift.llm.infer.protocol import ChatCompletionResponseChoice # noqa: E402
from swift.plugin import ORM, orms # noqa: E402
from swift.plugin.multi_turn import MultiTurnScheduler, multi_turns # noqa: E402
from swift.utils import get_logger # noqa: E402
logger = get_logger()
TOOL_CALL_RE = re.compile(r"<tool_call>\s*(.*?)\s*</tool_call>", re.DOTALL)
ANSWER_RE = re.compile(r"<answer>(.*?)</answer>", re.DOTALL)
_THINK_INNER = r"(?:(?!</think>|<tool_call>|<answer>).)*"
_TOOL_CALL_TURN_RE = re.compile(
rf"^<think>{_THINK_INNER}</think>\s*<tool_call>.*?</tool_call>(?![\s\S])",
re.DOTALL,
)
_ANSWER_TURN_RE = re.compile(
rf"^<think>{_THINK_INNER}</think>\s*<answer>.*?</answer>(?![\s\S])",
re.DOTALL,
)
def _is_canonical_format(completion: str) -> bool:
value = (completion or "").strip()
return bool(
_TOOL_CALL_TURN_RE.match(value)
or _ANSWER_TURN_RE.match(value)
)
def _batch_values(value: Any, count: int) -> List[Any]:
if isinstance(value, list) and len(value) == count:
return value
return [value] * count
class VRBenchNavigationReward(ORM):
"""Reward a final direction sequence that reaches the recorded goal."""
def __call__(
self,
completions,
task_id=None,
initial_state=None,
**kwargs,
) -> List[float]:
completions = list(completions or [])
tasks = _batch_values(task_id, len(completions))
states = _batch_values(initial_state, len(completions))
rewards: List[float] = []
for completion, task, state in zip(completions, tasks, states):
match = ANSWER_RE.search(completion or "")
if not match or task not in SUPPORTED_TASKS or not state:
rewards.append(0.0)
continue
actions = parse_action_tokens(match.group(1))
solved = bool(actions) and replay_actions(task, state, actions)
rewards.append(1.0 if solved else 0.0)
return rewards
orms["vrbench_navigation_reward"] = VRBenchNavigationReward
class VRBenchNavigationFormatReward(ORM):
"""Use the same canonical format reward scale as FrozenLake v3."""
def __call__(self, completions, **kwargs) -> List[float]:
rewards: List[float] = []
for completion in completions or []:
completion = completion or ""
match = ANSWER_RE.search(completion)
valid_actions = bool(
match and parse_action_tokens(match.group(1))
)
rewards.append(
0.3 * float(_is_canonical_format(completion))
+ 0.2 * float(valid_actions)
)
return rewards
orms["vrbench_navigation_format"] = VRBenchNavigationFormatReward
class VRBenchNavigationScheduler(MultiTurnScheduler):
"""FrozenLake-v3-style scheduler for three VRBench grid tasks."""
def __init__(self, max_turns: int = 100, *args, **kwargs):
super().__init__(max_turns=max_turns, *args, **kwargs)
self.tool_img_dir = Path(
os.environ.get(
"VRBENCH_NAV_IMG_DIR",
"/tmp/vrbench_navigation_rollout",
)
)
self.tool_img_dir.mkdir(parents=True, exist_ok=True)
def _state(self, infer_request: RolloutInferRequest) -> Dict[str, Any]:
if (
not hasattr(infer_request, "data_dict")
or infer_request.data_dict is None
):
infer_request.data_dict = {}
data = infer_request.data_dict
if "_vrbench_navigation" not in data:
task_id = data.get("task_id")
initial_state = data.get("initial_state")
if task_id not in SUPPORTED_TASKS:
raise RuntimeError(
"VRBenchNavigationScheduler requires task_id in "
f"{SUPPORTED_TASKS}; got {task_id!r}"
)
if not initial_state:
raise RuntimeError(
"VRBenchNavigationScheduler requires the dataset column "
"`initial_state`."
)
decoded = decode_initial_state(initial_state)
encoded = encode_initial_state(decoded)
data["initial_state"] = encoded
data["_vrbench_navigation"] = {
"task_id": task_id,
"initial_state": encoded,
"state": clone_state(decoded),
"done": False,
"success": False,
"actions": [],
}
if "rollout_uuid" not in data:
data["rollout_uuid"] = uuid.uuid4().hex[:12]
if "session_messages" not in data:
data["session_messages"] = [
{
"role": message["role"],
"content": message["content"],
}
for message in (infer_request.messages or [])
if message.get("role") != "assistant"
]
return data["_vrbench_navigation"]
@staticmethod
def _sample_id(infer_request: RolloutInferRequest) -> str:
data = infer_request.data_dict or {}
if "rollout_uuid" not in data:
data["rollout_uuid"] = uuid.uuid4().hex[:12]
infer_request.data_dict = data
return data["rollout_uuid"]
def check_finished(
self,
infer_request: RolloutInferRequest,
response_choice: ChatCompletionResponseChoice,
current_turn: int,
) -> bool:
state = self._state(infer_request)
content = response_choice.message.content or ""
if "<answer>" in content and "</answer>" in content:
return True
if state.get("done"):
return True
return super().check_finished(
infer_request,
response_choice,
current_turn,
)
def step(
self,
infer_request: RolloutInferRequest,
response_choice: ChatCompletionResponseChoice,
current_turn: int,
) -> Dict[str, Any]:
rollout_state = self._state(infer_request)
task_id = rollout_state["task_id"]
content = response_choice.message.content or ""
new_images: List[str] = []
match = TOOL_CALL_RE.search(content)
if not match:
body_text = (
"No <tool_call> found in your response. Call the move tool "
"with one direction, or finalize with "
"<answer>full action sequence space-separated</answer>."
)
else:
try:
tool_call = json.loads(match.group(1))
except json.JSONDecodeError:
body_text = (
"Your <tool_call> JSON was malformed. Use exactly:\n"
'<tool_call>{"name": "move", "arguments": '
'{"direction": "Down"}}</tool_call>'
)
else:
if not isinstance(tool_call, dict):
body_text = (
"Malformed tool call: the JSON value must be an object."
)
else:
name = tool_call.get("name")
arguments = tool_call.get("arguments") or {}
if name != "move":
body_text = (
f"Unknown tool {name!r}. Only `move` is available."
)
elif not isinstance(arguments, dict):
body_text = (
"Malformed move arguments: `arguments` must be "
"a JSON object."
)
else:
direction = arguments.get("direction", "")
outcome = step_state(
task_id,
rollout_state["state"],
direction,
)
body_text = outcome.body_text
if outcome.changed:
rollout_state["actions"].append(direction)
rollout_state["done"] = outcome.done
rollout_state["success"] = outcome.success
image_path = self.tool_img_dir / (
f"{self._sample_id(infer_request)}_"
f"t{current_turn:03d}.png"
)
try:
render_state(
task_id,
rollout_state["state"],
image_path,
)
except Exception as exc:
logger.warning(
f"VRBench navigation render failed: {exc}"
)
else:
new_images.append(str(image_path))
if rollout_state.get("done"):
body_text += (
"\n\nThe episode has ended. Output your full executed action "
"sequence as <answer>Direction1 Direction2 ...</answer>."
)
all_images = list(infer_request.images or []) + new_images
labelled = [
(
"initial state" if index == 0 else f"after step {index}",
path,
)
for index, path in enumerate(all_images)
]
continuation_prompt = create_tool_response(body_text, labelled)
data = infer_request.data_dict or {}
session_messages = data.get("session_messages")
if session_messages is not None:
session_messages.append(
{"role": "assistant", "content": content}
)
session_messages.append(
{"role": "user", "content": continuation_prompt}
)
next_messages = list(session_messages)
else:
next_messages = [
{"role": "user", "content": continuation_prompt}
]
new_request = RolloutInferRequest(
messages=next_messages,
images=all_images,
audios=getattr(infer_request, "audios", None),
videos=getattr(infer_request, "videos", None),
tools=getattr(infer_request, "tools", None),
objects=getattr(infer_request, "objects", None),
data_dict=infer_request.data_dict,
)
return {
"infer_request": new_request,
"rollout_infos": {
"task_id": task_id,
"navigation_actions": list(rollout_state["actions"]),
"navigation_done": rollout_state["done"],
"navigation_success": rollout_state["success"],
"num_turns": current_turn,
},
}
multi_turns["vrbench_navigation_scheduler"] = VRBenchNavigationScheduler
|