Datasets:
File size: 9,526 Bytes
8ad880b | 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 | """Assemble the unified response table: raw generations + quality (objective / judge / COMET / SARI+BERTScore) + measured cost.
Output: /mnt/localssd/seqroute1m/dataset/responses.parquet (one row per (query_id, model_id); never contains router-visible *features* of other responses)
Quality provenance per row: Q_source in {objective, judge, comet, sari_bertscore, missing}. `label_confidence` = 1 for deterministic scorers, the family ICC for judge scores.
Cost: gpu_seconds = rate_m * (completion_tokens + 0.1 * prompt_tokens) where rate_m = measured wave GPU-seconds / weighted tokens of the large waves (w1..w3) for model m
(smoke-wave GPU-seconds are not used for rates: 100-query batches leave the GPUs mostly idle). Latency is the per-request wall time measured under saturated batching.
"""
import json
import sys
from pathlib import Path
import numpy as np
import pandas as pd
ROOT = Path("/mnt/localssd/seqroute1m")
LP_COLS = ["mean_lp", "median_lp", "p05_lp", "p25_lp", "min_lp", "mean_lp_tail64", "frac_lp_lt_2", "n_lp", "mean_entropy_top5", "median_entropy_top5", "p90_entropy_top5", "max_entropy_top5"]
def judge_q(d):
"""Substantive dimensions (factuality, coverage) carry 80% of the judge score; relevance / coherence / adherence saturate near 10 and share the rest."""
return 0.4 * d.factuality + 0.4 * d.coverage + 0.2 * (d.relevance + d.coherence + d.instruction_adherence) / 3
def load_jsonl(path):
return pd.read_json(path, lines=True, dtype=False) if Path(path).exists() and Path(path).stat().st_size else pd.DataFrame()
def main():
models = json.load(open(ROOT / "manifests/models.json"))
mid = [m["id"] for m in models]
q = pd.read_parquet(ROOT / "queries/master_queries.parquet", columns=["query_id", "task_family", "source_dataset", "evaluator_type", "split", "in_20k"])
qfam = q.set_index("query_id").evaluator_type
parts = []
for m in mid:
f = ROOT / "responses" / f"{m}.jsonl"
if not f.exists():
continue
r = load_jsonl(f).rename(columns={"id": "query_id"})
r["model_id"] = m
parts.append(r)
R = pd.concat(parts, ignore_index=True)
R = R.drop_duplicates(["query_id", "model_id"], keep="first")
# ---- quality ------------------------------------------------------------------------------------------------------------------------
O = pd.concat([load_jsonl(ROOT / "scores/obj" / f"{m}.jsonl").assign(model_id=m) for m in mid if (ROOT / "scores/obj" / f"{m}.jsonl").exists()], ignore_index=True)
O = O.rename(columns={"Q": "Q", "raw": "raw_metric", "answer": "answer_extracted"})[["query_id", "model_id", "Q", "raw_metric", "raw_name", "parse_ok", "truncated", "empty_final", "answer_extracted", "sub"]]
O["Q_source"] = "objective"
aux = []
for name, src in (("comet.jsonl", "comet"), ("simplification.jsonl", "sari_bertscore")):
a = load_jsonl(ROOT / "scores/aux" / name)
if len(a):
a[["model_id", "query_id"]] = a["id"].str.split("|", n=1, expand=True)
a["Q_source"] = src
a = a.rename(columns={"raw": "raw_metric"})
for c in ("parse_ok", "truncated"):
a[c] = True if c == "parse_ok" else False
aux.append(a[["query_id", "model_id", "Q", "raw_metric", "raw_name", "empty_final", "Q_source", "parse_ok", "truncated"]])
J = load_jsonl(ROOT / "scores/judge/parsed_primary.jsonl")
jud = pd.DataFrame()
icc = {}
if len(J):
J = J[~J.id.str.startswith("ctrl_")]
J[["model_id", "query_id"]] = J["id"].str.split("|", n=1, expand=True)
J = J[J.rep == 0].copy()
J["Q_mean5"] = J.Q
J["Q"] = np.where(J.get("empty_final", False).fillna(False).astype(bool), 0.0, judge_q(J))
dims = [c for c in ["factuality", "coverage", "relevance", "coherence", "instruction_adherence"] if c in J.columns]
J["judge_dims"] = J[dims].apply(lambda r: json.dumps({k: (None if pd.isna(v) else float(v)) for k, v in r.items()}), axis=1) if dims else "{}"
jud = J.assign(raw_metric=J.get("Q"), raw_name="judge_mean5", Q_source="judge", parse_ok=J.parsed.astype(bool), truncated=False,
empty_final=J.get("empty_final", False).fillna(False).astype(bool))[["query_id", "model_id", "Q", "raw_metric", "raw_name", "Q_source", "parse_ok", "truncated", "empty_final", "judge_dims"]]
rep = load_jsonl(ROOT / "scores/judge/parsed_repeat.jsonl")
if len(rep):
rep = rep[rep.parsed & ~rep.id.str.startswith("ctrl_")]
rep[["model_id", "query_id"]] = rep["id"].str.split("|", n=1, expand=True)
rep["Q"] = judge_q(rep)
rep = rep.merge(q[["query_id", "task_family"]], on="query_id")
per_resp_sd = rep.groupby(["model_id", "query_id"]).Q.std(ddof=1).rename("judge_noise_sd").reset_index()
for fam, g in rep.groupby("task_family"):
within = g.groupby(["model_id", "query_id"]).Q.var(ddof=1).mean()
total = g.Q.var(ddof=1)
icc[fam] = dict(within_var=float(within), total_var=float(total), reliability=float(max(0.0, 1 - within / total)), pooled_noise_sd=float(np.sqrt(within)))
json.dump(icc, open(ROOT / "manifests/judge_reliability_icc.json", "w"), indent=1)
Qs = pd.concat([O] + aux + ([jud] if len(jud) else []), ignore_index=True)
Qs = Qs.drop_duplicates(["query_id", "model_id"], keep="first")
D = R.merge(Qs, on=["query_id", "model_id"], how="left").merge(q[["query_id", "task_family", "source_dataset", "split", "in_20k"]], on="query_id", how="left")
D["Q_source"] = D.Q_source.fillna("missing")
D["empty_final"] = D.empty_final.fillna(False).astype(bool)
# label confidence / judge noise
fam_noise = {f: v["pooled_noise_sd"] for f, v in icc.items()}
fam_rel = {f: v["reliability"] for f, v in icc.items()}
D["judge_noise_sd"] = np.where(D.Q_source == "judge", D.task_family.map(fam_noise), 0.0)
D["label_confidence"] = np.where(D.Q_source == "judge", D.task_family.map(fam_rel).fillna(0.0), np.where(D.Q_source == "missing", 0.0, 1.0))
D.loc[(D.Q_source == "judge") & D.empty_final, ["judge_noise_sd", "label_confidence"]] = [0.0, 1.0]
# ---- cost -----------------------------------------------------------------------------------------------------------------------------
D["wtok"] = D.completion_tokens + 0.1 * D.prompt_tokens
waves = {w: set(json.load(open(ROOT / "manifests" / f"wave_{w}.json"))) for w in ("w1", "w2", "w3") if (ROOT / "manifests" / f"wave_{w}.json").exists()}
rates = {}
for m in mid:
gs, wt = 0.0, 0.0
for w, ids in waves.items():
sf = ROOT / "responses" / f"{m}.{w}.stats.json"
if not sf.exists():
continue
st = json.load(open(sf))
if st["n_generated"] < 500:
continue
gs += st["gpu_seconds"]
wt += st["completion_tokens"] + 0.1 * st["prompt_tokens"]
rates[m] = gs / wt if wt else np.nan
def ts_rate(m):
"""Fallback when wave stats were overwritten by retry rounds: 8 GPUs x (0.1%-trimmed) wave window from response timestamps; capped attempts archived by the extended pass are included, regenerated ones are not."""
arch = ROOT / "responses_truncated" / f"{m}.jsonl"
ar = [json.loads(l) for l in open(arch)] if arch.exists() else []
aid = {r["id"] for r in ar}
main_rows = [json.loads(l) for l in open(ROOT / "responses" / f"{m}.jsonl")]
smoke = set(json.load(open(ROOT / "manifests/wave_smoke.json")))
gs_, wt_ = 0.0, 0.0
for w, ids in waves.items():
rows = [r for r in main_rows if r["id"] in ids and r["id"] not in aid and r["id"] not in smoke] + [r for r in ar if r["id"] in ids and r["id"] not in smoke]
gs_ += 8 * (np.quantile([r["t_end"] for r in rows], 0.999) - min(r["t_end"] - r["latency_s"] for r in rows))
wt_ += sum(r["completion_tokens"] + 0.1 * r["prompt_tokens"] for r in rows)
return gs_ / wt_
ts_fallback = [m for m in mid if np.isnan(rates[m])]
for m in ts_fallback:
rates[m] = ts_rate(m)
if all(np.isnan(v) for v in rates.values()): # smoke-only run: fall back to smoke stats (flagged)
for m in mid:
sf = ROOT / "responses" / f"{m}.smoke.stats.json"
if sf.exists():
st = json.load(open(sf))
rates[m] = st["gpu_seconds"] / (st["completion_tokens"] + 0.1 * st["prompt_tokens"])
D["cost_source"] = "smoke_wave_rate(NOT FINAL)"
else:
D["cost_source"] = "wave_calibrated_token_rate"
D["gpu_seconds"] = D.wtok * D.model_id.map(rates)
ref = D[D.split == "train"].groupby("task_family").gpu_seconds.median()
D["cost_norm"] = D.gpu_seconds / D.task_family.map(ref)
D = D.drop(columns=["wtok"])
for c in LP_COLS:
if c not in D.columns:
D[c] = np.nan
out = ROOT / "dataset"
out.mkdir(exist_ok=True)
D.to_parquet(out / "responses.parquet", compression="zstd", index=False)
json.dump(dict(rate_from_timestamps_for=ts_fallback, gpu_seconds_per_weighted_token=rates, reference_cost_median_gpu_s_by_family=ref.to_dict()), open(ROOT / "manifests/cost_calibration.json", "w"), indent=1)
print("responses", len(D), "missing quality", int((D.Q_source == "missing").sum()), "models", D.model_id.nunique(), "queries", D.query_id.nunique())
if __name__ == "__main__":
main()
|