File size: 4,888 Bytes
a91390f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 | import json
import re
from pathlib import Path
import pandas as pd
ROOT = Path("completed_experiments/evals/full-2026-10-23/outputs")
CORPUS = Path("questions/corpus.jsonl")
OUT = Path("analysis/designed-difficulty")
MODELS = [
"qwen3.5-4b",
"gemma-4-e4b",
"granite-4.2-3b",
"nemotron-3-nano-4b",
"k2-horizon-3.7b",
]
PATTERN = re.compile(
r"^(?P<variant>[^/]+)/(?P<question_id>[^#]+)"
r"#e(?P<epoch>\d+)#t(?P<turn>\d+)$"
)
# 1. Load question metadata
questions = {}
with CORPUS.open(encoding="utf-8") as f:
for line in f:
if not line.strip():
continue
q = json.loads(line)
qid = q["id"]
if qid in questions:
raise ValueError(f"Duplicate question ID: {qid}")
difficulty = q.get("difficulty") or {}
questions[qid] = {
"claim_id": q.get("claim_id"),
"category": q.get("category"),
"framing": q.get("framing"),
"evidence_scarcity": difficulty.get("evidence_scarcity"),
"question_openness": difficulty.get("question_openness"),
"band": difficulty.get("band"),
}
print("Corpus questions:", len(questions))
# 2. Load judge results
rows = []
unparsed = []
for model in MODELS:
path = ROOT / model / "pplx/results.jsonl"
if not path.exists():
raise FileNotFoundError(path)
count = 0
with path.open(encoding="utf-8") as f:
for line_no, line in enumerate(f, 1):
if not line.strip():
continue
result = json.loads(line)
case_id = result.get("case_id", "")
match = PATTERN.fullmatch(case_id)
if not match:
unparsed.append((model, line_no, case_id))
continue
parts = match.groupdict()
qid = parts["question_id"]
meta = questions.get(qid, {})
probabilities = result.get("probabilities") or {}
rows.append({
"model": model,
"variant": parts["variant"],
"question_id": qid,
"epoch": int(parts["epoch"]),
"turn": int(parts["turn"]),
"case_id": case_id,
"choice": result.get("choice"),
"status": result.get("status"),
"dropped": result.get("dropped"),
"confidence": result.get("confidence"),
"p_hold": probabilities.get("hold"),
"p_diverge": probabilities.get("diverge"),
"p_abstain": probabilities.get("abstain"),
**meta,
})
count += 1
print(f"{model}: {count:,} records")
df = pd.DataFrame(rows)
if df.empty:
raise RuntimeError("No parsed Judge records found")
# 3. Integrity checks
key_cols = ["model", "case_id"]
duplicates = df.duplicated(key_cols, keep=False)
missing_metadata = df["evidence_scarcity"].isna() | df["question_openness"].isna()
valid_labels = {"hold", "diverge", "abstain"}
df["valid"] = (
df["status"].eq("ok")
& df["dropped"].eq(0)
& df["choice"].isin(valid_labels)
)
df["diverged"] = df["choice"].eq("diverge") & df["valid"]
print("\n=== INTEGRITY CHECK ===")
print("Total records:", len(df))
print("Unparsed case IDs:", len(unparsed))
print("Duplicate model/case IDs:", int(duplicates.sum()))
print("Missing difficulty metadata:", int(missing_metadata.sum()))
print("Valid Judge records:", int(df["valid"].sum()))
print("\nStatus distribution:")
print(df["status"].value_counts(dropna=False).to_string())
print("\nChoice distribution:")
print(df["choice"].value_counts(dropna=False).to_string())
print("\nVariant distribution:")
print(pd.crosstab(df["model"], df["variant"]).to_string())
if unparsed:
print("\nUnparsed examples:", unparsed[:10])
# 4. Save the raw joined dataset
OUT.mkdir(parents=True, exist_ok=True)
output = OUT / "baseline_joined.csv"
df.to_csv(output, index=False)
print(f"\nSaved: {output}")
# 5. Preliminary summary (only if integrity checks pass)
if duplicates.any() or missing_metadata.any() or unparsed:
print("\nWARNING: Resolve integrity issues before computing rates.")
else:
opening = df[(df["variant"] == "multiturn") & (df["turn"] == 0)]
print("\n=== MULTITURN OPENING: DIVERGENCE RATE ===")
print("Denominator: valid hold/diverge/abstain judgments")
summary = (
opening[opening["valid"]]
.groupby("model")
.agg(
n=("diverged", "size"),
questions=("question_id", "nunique"),
diverged=("diverged", "sum"),
rate=("diverged", "mean"),
)
.reset_index()
)
summary["rate_pct"] = (summary["rate"] * 100).round(2)
print(
summary[
["model", "questions", "n", "diverged", "rate_pct"]
].to_string(index=False)
)
|