| """检查点状态：GRPO 的"题目"。 |
| |
| CheckpointState = curator 在一个检查点看到的全部输入（纯文本，可序列化）： |
| 技能库快照 + episode 批 + Reflector 分析 + 上轮回执 + 工具规格。 |
| 同一 state 采 G 条 op 输出 → 组内比较 → 这就是 GRPO 的一组。 |
| |
| 状态池两个来源： |
| 1) 离线收集（冷启动）：training-free 跑（run_evolve）的每个检查点都是一个天然状态 |
| —— collect_from_run() 从 runs/*/episodes.jsonl + snapshots/ 里回收； |
| 2) 在线生成（正式训练）：SessionSpec 驱动 rollout 引擎，边跑边产状态（TODO Session-B）。 |
| """ |
|
|
| import json |
| import os |
| import glob |
| import sys |
| from dataclasses import dataclass, field, asdict |
| from typing import List, Optional |
|
|
| sys.path.insert(0, "/mnt/tokyo-ai/gst/EvoStream") |
|
|
|
|
| @dataclass |
| class CheckpointState: |
| state_id: str |
| dataset: str |
| lib_snapshot: dict |
| episodes: List[dict] |
| analysis: Optional[dict] |
| last_receipt: str |
| followup_qids: List[str] = field(default_factory=list) |
| gt_map: dict = field(default_factory=dict) |
|
|
| def to_json(self): |
| return json.dumps(asdict(self), ensure_ascii=False) |
|
|
| @staticmethod |
| def from_json(s: str): |
| return CheckpointState(**json.loads(s)) |
|
|
|
|
| def build_curator_prompt(state: CheckpointState, cfg) -> str: |
| """复用推理侧的 CURATOR_PROMPT 模板拼训练输入（保证 prompted/训练版分布一致）。""" |
| from infer.curator.curator import CURATOR_PROMPT, PARAM_RULE_ON |
| from infer.curator.reflector import format_episodes |
| from infer.tools import TOOL_SPECS |
|
|
| cards = state.lib_snapshot.get("cards", []) |
| catalog_lines = [] |
| for c in cards: |
| m = c.get("meta_info", {}) |
| catalog_lines.append( |
| f"- [{c['card_id']}] inj={m.get('injected',0)} fired={m.get('fired',0)} " |
| f"helped={m.get('helped',0)} harmful={m.get('harmful',0)}\n" |
| f" When: {c['trigger'][:150]}\n Do: {c['program'][:200]}") |
| tool_specs = "\n".join( |
| f"- {name}: {spec['description']}" for name, spec in TOOL_SPECS.items()) |
|
|
| if state.analysis is not None: |
| analysis_block = json.dumps( |
| {k: state.analysis.get(k) for k in ("failure_patterns", "observations", "summary")}, |
| ensure_ascii=False, indent=1)[:4000] |
| else: |
| analysis_block = "## Raw episodes\n" + format_episodes(state.episodes)[:5000] |
|
|
| return CURATOR_PROMPT.format( |
| tool_specs=tool_specs, |
| version=state.lib_snapshot.get("version", 0), n_cards=len(cards), |
| skill_catalog="\n".join(catalog_lines) or "(empty library)", |
| last_receipt=state.last_receipt, analysis_block=analysis_block, |
| max_ops=3, param_rule=PARAM_RULE_ON) |
|
|
|
|
| def collect_from_run(run_dir: str, out_dir: str, dataset: str = "ovo") -> int: |
| """从 training-free run 回收检查点状态（冷启动状态池）。 |
| |
| 近似拼装：第 k 个检查点的 lib 快照 = snapshots/skills_v{k-1}（策展前状态）， |
| episodes = 该检查点消化的批次（按 skill_lib_version 分组）。 |
| """ |
| os.makedirs(out_dir, exist_ok=True) |
| eps = [json.loads(l) for l in open(os.path.join(run_dir, "episodes.jsonl"))] |
| snaps = sorted(glob.glob(os.path.join(run_dir, "snapshots", "skills_v*.json"))) |
| if not snaps: |
| return 0 |
| by_ver = {} |
| for e in eps: |
| by_ver.setdefault(e.get("skill_lib_version", 0), []).append(e) |
|
|
| n = 0 |
| run_name = os.path.basename(run_dir.rstrip("/")) |
| for snap in snaps: |
| ver = int(snap.rsplit("skills_v", 1)[1].split(".")[0]) |
| batch = by_ver.get(ver - 1) or by_ver.get(ver) |
| if not batch: |
| continue |
| lib = json.load(open(snap)) |
| for c in lib.get("cards", []): |
| c.pop("_trigger_emb", None) |
| st = CheckpointState( |
| state_id=f"{run_name}_ckpt{ver}", dataset=dataset, |
| lib_snapshot=lib, episodes=batch, analysis=None, |
| last_receipt="(collected offline)") |
| with open(os.path.join(out_dir, f"{st.state_id}.json"), "w") as f: |
| f.write(st.to_json()) |
| n += 1 |
| return n |
|
|
|
|
| def load_state_pool(pool_dir: str) -> List[CheckpointState]: |
| return [CheckpointState.from_json(open(p).read()) |
| for p in sorted(glob.glob(os.path.join(pool_dir, "*.json")))] |
|
|
