step_train/src/evaluation/eval_metrics.py

"""Eval metric aggregation for MetaMem on LongMemEval-S.

Given a list of per-query eval records (one per query), compute:
  - answer accuracy: EM / F1 / judge, overall + by FIR state (ms_label) + by query_type
  - retrieval: recall@k / MRR (non-DIRECT only)
  - strategy: over-retrieval rate, under-retrieval rate, strategy accuracy (vs oracle)
  - hallucinated-entity rate
  - cost: avg retrieval calls, retrieved tokens, generated tokens

Each record dict is expected to have:
  query_id, user_id, query_type, oracle_ms (or None), pred_ms, act,
  answer_em (bool), answer_f1 (float), answer_judge (bool|None),
  retrieved (list of session_ids), gold_evidence_ids (list),
  new_rank (int|None), retrieved_tokens, retrieval_calls, gen_tokens,
  hallucinated (bool)
"""

from collections import defaultdict
from typing import Dict, List, Optional


def _mean(xs):
    xs = [x for x in xs if x is not None]
    return (sum(xs) / len(xs)) if xs else 0.0


def aggregate(records: List[dict], k: int = 10) -> dict:
    out: Dict = {}

    # ---- answer accuracy ----
    out["n"] = len(records)
    out["em"] = _mean([r["answer_em"] for r in records])
    out["f1"] = _mean([r["answer_f1"] for r in records])
    # answer_judge may be the three-tier verdict string (correct/partial/wrong, new LLMJudge)
    # or a legacy bool. Map to numeric: correct=1.0, partial=0.5, wrong=0.0, True/False=1/0.
    def _judge_num(v):
        if isinstance(v, str):
            return {"correct": 1.0, "partial": 0.5, "wrong": 0.0}.get(v)
        if isinstance(v, bool):
            return 1.0 if v else 0.0
        return None
    judges = [_judge_num(r.get("answer_judge")) for r in records]
    judges = [j for j in judges if j is not None]
    out["judge_acc"] = _mean(judges) if judges else None
    # strict-correct rate (verdict=="correct" only), excludes partial-credit
    jc = [1.0 if r.get("answer_judge") == "correct" else 0.0
          for r in records if r.get("answer_judge") is not None]
    out["judge_correct_rate"] = round(_mean(jc), 4) if jc else None

    def grouped(key, valfn):
        g = defaultdict(list)
        for r in records:
            g[r.get(key)].append(valfn(r))
        return {str(kk): round(_mean(vv), 4) for kk, vv in g.items()}

    out["em_by_ms"] = grouped("oracle_ms", lambda r: r["answer_em"])
    out["f1_by_ms"] = grouped("oracle_ms", lambda r: r["answer_f1"])
    if judges:
        out["judge_by_ms"] = grouped("oracle_ms",
                                     lambda r: _judge_num(r.get("answer_judge")) or 0.0)
        out["judge_by_qtype"] = grouped("query_type",
                                        lambda r: _judge_num(r.get("answer_judge")) or 0.0)
    out["em_by_qtype"] = grouped("query_type", lambda r: r["answer_em"])

    # ---- retrieval recall@k / MRR (non-DIRECT only) ----
    non_direct = [r for r in records if r["act"] != "DIRECT"]
    recalls, rrs = [], []
    for r in non_direct:
        gold = set(r.get("gold_evidence_ids", []))
        retrieved = r.get("retrieved", [])[:k]
        hit = bool(set(retrieved) & gold)
        recalls.append(1.0 if hit else 0.0)
        rr = 0.0
        for rank, rid in enumerate(retrieved, 1):
            if rid in gold:
                rr = 1.0 / rank
                break
        rrs.append(rr)
    out["recall_at_k"] = round(_mean(recalls), 4) if recalls else None
    out["mrr"] = round(_mean(rrs), 4) if rrs else None
    out["k"] = k

    # ---- strategy: over/under-retrieval + strategy accuracy ----
    # should_search = oracle != SM (only meaningful when oracle present). On LongMemEval-S
    # only the FIR-labelled subset has oracle_ms; surface the subset size so these rates
    # are never misread as a full-500-user metric.
    with_oracle = [r for r in records if r.get("oracle_ms")]
    out["n_with_oracle"] = len(with_oracle)
    out["n_without_oracle"] = out["n"] - len(with_oracle)
    over, under, correct = [], [], []
    for r in with_oracle:
        should = r["oracle_ms"] != "SM"
        searched = r["act"] != "DIRECT"
        over.append(1.0 if (not should and searched) else 0.0)
        under.append(1.0 if (should and not searched) else 0.0)
        correct.append(1.0 if (should == searched) else 0.0)
    out["over_retrieval_rate"] = round(_mean(over), 4) if over else None
    out["under_retrieval_rate"] = round(_mean(under), 4) if under else None
    out["strategy_accuracy"] = round(_mean(correct), 4) if correct else None

    # ---- hallucination (only meaningful for NM/VM non-DIRECT, matching reward) ----
    hsubset = [r for r in records if r.get("oracle_ms") in ("NM", "VM") and r.get("act") != "DIRECT"]
    out["hallucinated_rate"] = round(_mean([1.0 if r.get("hallucinated") else 0.0 for r in hsubset]), 4) if hsubset else None
    out["hallucinated_subset_n"] = len(hsubset)

    # ---- cost ----
    out["avg_retrieval_calls"] = round(_mean([r.get("retrieval_calls", 0) for r in records]), 3)
    out["avg_retrieved_tokens"] = round(_mean([r.get("retrieved_tokens", 0) for r in records]), 1)
    out["avg_gen_tokens"] = round(_mean([r.get("gen_tokens", 0) for r in records]), 1)

    # ---- action distribution ----
    act_counts = defaultdict(int)
    for r in records:
        act_counts[r["act"]] += 1
    out["action_distribution"] = dict(act_counts)
    return out