| """Step 10a: Precompute BGE-M3 top-10 window base_rank for DAPO anchors. |
| |
| For each query that has an ms_label file (MAJORITY + AMBIGUOUS merged), build the user's |
| BGE-M3 user-turn-only index and compute get_rank(original_query, gold_evidence_ids) |
| within the top-10 window (-1 if outside). Export anchors.jsonl with oracle_label, |
| cartridge_path, gold_evidence_ids, gold_aliases, base_rank (+ full_rank diag). |
| oracle_label is a passive diagnostic field only — DAPO does NOT train on it (the sampler |
| is uniform and label-agnostic; the reward is purely outcome-driven). |
| |
| Requires GPU + FlagEmbedding (BGE-M3). |
| |
| Usage: |
| python scripts/train/10a_precompute_base_rank.py |
| """ |
|
|
| import argparse |
| import json |
| import os |
| import sys |
|
|
| sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))) |
| PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) |
|
|
| from src.retrieval.bge_m3_retriever import BGEM3Retriever |
| from src.utils import load_json, load_yaml, setup_logger |
| from src.utils.cartridge_utils import find_cartridge_path |
|
|
| logger = setup_logger(__name__) |
|
|
|
|
| def _r(p): |
| return p if os.path.isabs(p) else os.path.join(PROJECT_ROOT, p) |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--dataset", default="data/processed/metamem/dataset.json") |
| ap.add_argument("--ms-dir", default="data/labeled/ms_labels/Qwen3-8B") |
| ap.add_argument("--cartridge-dir", default="checkpoints/cartridge/Qwen3-8B") |
| ap.add_argument("--retrieval-config", default="configs/retrieval/bge_m3.yaml") |
| ap.add_argument("--out", default="data/train/rl/anchors.jsonl") |
| ap.add_argument("--include-query-ids", default=None, |
| help="JSON list of query_ids to KEEP (e.g. splits/rl_ids.json). " |
| "Omit to use all labeled queries (legacy behavior).") |
| args = ap.parse_args() |
|
|
| dataset = load_json(_r(args.dataset)) |
| ms_dir = _r(args.ms_dir) |
| cartridge_dir = _r(args.cartridge_dir) |
| rcfg = load_yaml(_r(args.retrieval_config)) |
| out_path = _r(args.out) |
| os.makedirs(os.path.dirname(out_path), exist_ok=True) |
|
|
| include_ids = None |
| if args.include_query_ids: |
| include_ids = set(load_json(_r(args.include_query_ids))) |
| logger.info(f"Filtering to {len(include_ids)} query_ids from {args.include_query_ids}") |
|
|
| retriever = BGEM3Retriever(rcfg) |
|
|
| anchors = [] |
| n_total = 0 |
| n_no_cart = 0 |
| for ud in dataset: |
| uid = ud["user_id"] |
| |
| if include_ids is not None and not any(q["id"] in include_ids for q in ud["queries"]): |
| continue |
| cpath = find_cartridge_path(cartridge_dir, uid) |
| if cpath is None: |
| n_no_cart += 1 |
| continue |
| |
| retriever.build_user_index(uid, ud["user_sessions"]) |
| for q in ud["queries"]: |
| qid = q["id"] |
| if include_ids is not None and qid not in include_ids: |
| continue |
| n_total += 1 |
| ms_path = os.path.join(ms_dir, uid, f"{qid}_label.json") |
| if not os.path.exists(ms_path): |
| continue |
| ms = load_json(ms_path) |
| oracle = ms["ms_label"] |
| evidence_list = q.get("evidence_session_ids", []) |
| evidence = set(evidence_list) |
| base_rank = retriever.get_rank(uid, q["query"], evidence, window=retriever.default_top_k) |
| full_rank = retriever.full_rank(uid, q["query"], evidence) |
| extra = q.get("extra") or {} |
| qdate = extra.get("question_date") or extra.get("unified_date") |
| anchors.append({ |
| "user_id": uid, |
| "query_id": qid, |
| "query": q["query"], |
| "gold_answer": q["answer"], |
| "gold_aliases": [q["answer"]], |
| "gold_evidence_ids": list(evidence_list), |
| "oracle_label": oracle, |
| "cartridge_path": cpath, |
| "base_rank": base_rank, |
| "full_rank": full_rank, |
| "question_date": qdate, |
| }) |
|
|
| with open(out_path, "w", encoding="utf-8") as f: |
| for a in anchors: |
| f.write(json.dumps(a, ensure_ascii=False) + "\n") |
|
|
| from collections import Counter |
| by_ms = Counter(a["oracle_label"] for a in anchors) |
| in_window = sum(1 for a in anchors if a["base_rank"] != -1) |
| logger.info(f"Exported {len(anchors)} anchors → {out_path}") |
| logger.info(f" by oracle MS: {dict(by_ms)}") |
| logger.info(f" base_rank in top-10 window: {in_window}/{len(anchors)}") |
| logger.info(f" total queries seen {n_total}, users w/o cartridge {n_no_cart}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
