File size: 4,530 Bytes
0c64c71
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Judge validation: controls, repeat noise, cross-judge agreement, agreement with reference-based ROUGE-L.  Writes manifests/judge_validation.json."""
import json
import sys
from pathlib import Path

import numpy as np
import pandas as pd
from scipy.stats import pearsonr, spearmanr

sys.path.insert(0, "src")
ROOT = Path("/mnt/localssd/seqroute1m")
S = ROOT / "scores/judge"


def jq(d):
    return 0.4 * d.factuality + 0.4 * d.coverage + 0.2 * (d.relevance + d.coherence + d.instruction_adherence) / 3


def load(name):
    d = pd.read_json(S / f"parsed_{name}.jsonl", lines=True)
    return d


q = pd.read_parquet(ROOT / "queries/queries_20k.parquet", columns=["query_id", "task_family", "reference", "prompt"]).set_index("query_id")
out = {}
P = load("primary")
raw = pd.read_json(S / "raw_primary.jsonl", lines=True)
out["primary_calls"] = int(raw.id.str.contains("#r0$").sum())
out["primary_retried_calls"] = int(raw.id.str.contains("~rt").sum())
out["primary_unparsed"] = int((~P.parsed).sum())
P = P[P.parsed & P.get("empty_final", pd.Series(False, index=P.index)).fillna(False).eq(False)].copy()
P["Qw"] = jq(P)
ctrl = P[P.id.str.startswith("ctrl_")].copy()
ctrl["kind"] = ctrl.id.str.split("|").str[0]
ctrl["query_id"] = ctrl.id.str.split("|").str[1]
ctrl["family"] = ctrl.query_id.map(q.task_family)
out["controls_mean_Q"] = ctrl.groupby(["family", "kind"]).Qw.mean().round(3).unstack().to_dict()
pos, neg = ctrl[ctrl.kind == "ctrl_ref"].set_index("query_id").Qw, ctrl[ctrl.kind == "ctrl_neg"].set_index("query_id").Qw
both = pd.concat([pos.rename("ref"), neg.rename("neg")], axis=1).dropna()
out["control_pairs"] = int(len(both))
out["control_ref_gt_neg_rate"] = float((both.ref > both.neg).mean())
out["control_neg_Q_below_3_rate"] = float((both.neg < 3).mean())
M = P[~P.id.str.startswith("ctrl_")].copy()
M[["model_id", "query_id"]] = M.id.str.split("|", n=1, expand=True)
M["family"] = M.query_id.map(q.task_family)
out["Q_saturation"] = {f: dict(mean=float(g.Qw.mean()), share_ge_9=float((g.Qw >= 9).mean()), share_eq_10=float((g.Qw >= 9.99).mean())) for f, g in M.groupby("family")}
# agreement with ROUGE-L against the dataset reference summary
try:
    from rouge_score import rouge_scorer
    sc = rouge_scorer.RougeScorer(["rougeL"])
    resp = {}
    for mfile in (ROOT / "responses").glob("*.jsonl"):
        pass
    sub = M.sample(min(3000, len(M)), random_state=0)
    vals = []
    cache = {}
    for r in sub.itertuples():
        if r.model_id not in cache:
            cache[r.model_id] = {j["id"]: j for j in map(json.loads, open(ROOT / "responses" / f"{r.model_id}.jsonl"))}
        t = cache[r.model_id][r.query_id]["text"]
        t = t.rsplit("</think>", 1)[-1].strip()
        vals.append(sc.score(q.reference[r.query_id], t)["rougeL"].fmeasure)
    sub = sub.assign(rougeL=vals)
    out["corr_with_rougeL"] = {f: dict(pearson=float(pearsonr(g.Qw, g.rougeL)[0]), spearman=float(spearmanr(g.Qw, g.rougeL)[0]), n=int(len(g))) for f, g in sub.groupby("family")}
except Exception as e:
    out["corr_with_rougeL"] = f"unavailable: {e!r}"
# cross-judge
try:
    A = load("alt")
    A = A[A.parsed & ~A.id.str.startswith("ctrl_")].copy()
    A["Qw"] = jq(A)
    A[["model_id", "query_id"]] = A.id.str.split("|", n=1, expand=True)
    J = M.merge(A[["id", "Qw"]], on="id", suffixes=("", "_alt"))
    out["cross_judge"] = {f: dict(n=int(len(g)), pearson=float(pearsonr(g.Qw, g.Qw_alt)[0]), spearman=float(spearmanr(g.Qw, g.Qw_alt)[0]), mean_primary=float(g.Qw.mean()), mean_alt=float(g.Qw_alt.mean()),
                                  mean_abs_diff=float((g.Qw - g.Qw_alt).abs().mean())) for f, g in J.groupby("family")}
    mp = J.groupby("model_id").Qw.mean(); ma = J.groupby("model_id").Qw_alt.mean()
    out["cross_judge_model_ranking_spearman"] = float(spearmanr(mp, ma.reindex(mp.index))[0])
except Exception as e:
    out["cross_judge"] = f"unavailable: {e!r}"
# repeat noise on the weighted score
try:
    R = load("repeat"); R = R[R.parsed & ~R.id.str.startswith("ctrl_")].copy(); R["Qw"] = jq(R)
    R["family"] = R.id.str.split("|").str[1].map(q.task_family)
    out["repeat"] = {f: dict(n_responses=int(g.id.nunique()), within_sd=float(np.sqrt(g.groupby("id").Qw.var(ddof=1).mean())), total_sd=float(g.Qw.std(ddof=1))) for f, g in R.groupby("family")}
except Exception as e:
    out["repeat"] = f"unavailable: {e!r}"
json.dump(out, open(ROOT / "manifests/judge_validation.json", "w"), indent=1)
print(json.dumps({k: out[k] for k in out if k not in ("Q_saturation",)}, indent=1)[:3000])