Spaces:
Sleeping
Sleeping
File size: 11,364 Bytes
f39c634 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 | """Per-component capability analysis: base vs trained, from inference JSONs.
Mines the inference output JSONs in eval/results/ (which carry per-scenario
`components` + per-step `trajectory` with ground-truth reversibility labels)
to produce direct evidence of *what skill the model learned* — not just the
aggregate reward delta.
For a given model prefix it reads the four phase files:
<prefix>_base_t12.json <prefix>_base_t34.json
<prefix>_train_t12.json <prefix>_train_t34.json
combines t12+t34 into a base set and a trained set, and reports:
1. Per-component reward delta (the 6 rubric components)
2. Per-tier reward + reversibility_correct delta
3. Action-level reversibility PREDICTION ACCURACY (overall / on-irreversible /
on-reversible) — the load-bearing "did it learn to reason about
irreversibility" number
4. Confidence calibration: mean confidence when the prediction was right vs
wrong (overconfidence-on-errors should drop after training)
5. Action-type distribution shift
Writes a markdown report to eval/results/<prefix>_capability_report.md.
Usage:
python eval/capability_report.py --prefix qwen7b
python eval/capability_report.py --prefix llama8b
python eval/capability_report.py --prefix qwen7b --results-dir eval/results
"""
from __future__ import annotations
import argparse
import json
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any
# The 6 rubric reward components, in weight order, with their stored keys.
COMPONENTS = [
("viveka.reversibility_correct", "reversibility_correct", 0.30),
("viveka.task_progress", "task_progress", 0.25),
("viveka.confirmation_appropriate", "confirmation_appropriate", 0.15),
("viveka.confidence_brier", "confidence_brier", 0.15),
("viveka.over_asking", "over_asking", 0.10),
("viveka.hallucination", "hallucination", 0.05),
]
def _load_scenarios(results_dir: Path, prefix: str, split: str) -> list[dict[str, Any]]:
"""Load and concatenate the t12 + t34 scenario lists for a base/trained split."""
scenarios: list[dict[str, Any]] = []
for tiers in ("t12", "t34"):
path = results_dir / f"{prefix}_{split}_{tiers}.json"
if not path.exists():
print(f" [warn] missing {path}")
continue
data = json.loads(path.read_text())
scenarios.extend(data.get("scenarios", []))
return scenarios
def _mean(xs: list[float]) -> float | None:
xs = [x for x in xs if x is not None]
return sum(xs) / len(xs) if xs else None
def _fmt(v: float | None, prec: int = 4) -> str:
return f"{v:.{prec}f}" if isinstance(v, (int, float)) else "n/a"
def _component_means(scenarios: list[dict]) -> dict[str, float | None]:
out: dict[str, float | None] = {}
for key, _short, _w in COMPONENTS:
out[key] = _mean([s.get("components", {}).get(key) for s in scenarios])
return out
def _per_tier(scenarios: list[dict]) -> dict[int, dict[str, float | None]]:
by_tier: dict[int, list[dict]] = defaultdict(list)
for s in scenarios:
by_tier[int(s.get("tier_id", -1))].append(s)
out: dict[int, dict[str, float | None]] = {}
for tier, scen in sorted(by_tier.items()):
out[tier] = {
"n": len(scen),
"reward": _mean([s.get("reward") for s in scen]),
"reversibility_correct": _mean(
[s.get("components", {}).get("viveka.reversibility_correct") for s in scen]
),
}
return out
def _reversibility_accuracy(scenarios: list[dict]) -> dict[str, Any]:
"""Action-level: how often did predicted_reversibility match ground truth?
Splits by ground-truth class because getting IRREVERSIBLE right is the
safety-critical case (false 'reversible' on an irreversible op = the failure
mode the whole env is about).
"""
overall_correct, overall_total = 0, 0
irr_correct, irr_total = 0, 0
rev_correct, rev_total = 0, 0
conf_when_right: list[float] = []
conf_when_wrong: list[float] = []
for s in scenarios:
for step in s.get("trajectory", []):
gt = step.get("ground_truth_reversibility")
pred = step.get("predicted_reversibility")
if gt is None or pred is None:
continue
correct = step.get("correctness")
if correct is None:
correct = int(pred == gt)
overall_total += 1
overall_correct += int(bool(correct))
# ground-truth class split
if "irreversible" in str(gt):
irr_total += 1
irr_correct += int(bool(correct))
else:
rev_total += 1
rev_correct += int(bool(correct))
# calibration: confidence on right vs wrong predictions
conf = step.get("confidence")
if isinstance(conf, (int, float)):
(conf_when_right if correct else conf_when_wrong).append(float(conf))
def rate(c: int, t: int) -> float | None:
return c / t if t else None
return {
"overall_acc": rate(overall_correct, overall_total),
"overall_n": overall_total,
"irreversible_acc": rate(irr_correct, irr_total),
"irreversible_n": irr_total,
"reversible_acc": rate(rev_correct, rev_total),
"reversible_n": rev_total,
"conf_when_right": _mean(conf_when_right),
"conf_when_wrong": _mean(conf_when_wrong),
}
def _action_types(scenarios: list[dict]) -> Counter:
c: Counter = Counter()
for s in scenarios:
for step in s.get("trajectory", []):
at = step.get("action_type")
if at:
c[at] += 1
return c
def _delta_str(base: float | None, trained: float | None) -> str:
if base is None or trained is None:
return "n/a"
d = trained - base
return f"{d:+.4f}"
def build_report(prefix: str, results_dir: Path) -> str:
base = _load_scenarios(results_dir, prefix, "base")
trained = _load_scenarios(results_dir, prefix, "train")
if not base or not trained:
raise SystemExit(
f"Could not load base/trained scenarios for prefix {prefix!r} in {results_dir}. "
f"Expected files like {prefix}_base_t12.json, {prefix}_train_t34.json."
)
bc = _component_means(base)
tc = _component_means(trained)
bt = _per_tier(base)
tt = _per_tier(trained)
ba = _reversibility_accuracy(base)
ta = _reversibility_accuracy(trained)
bact = _action_types(base)
tact = _action_types(trained)
L: list[str] = []
L.append(f"# Capability Report — {prefix}")
L.append("")
L.append(f"Base: {len(base)} scenarios | Trained: {len(trained)} scenarios "
f"(t12 + t34 combined). Source: inference JSONs in `{results_dir}/`.")
L.append("")
# 1. Per-component reward delta
L.append("## 1. Per-component reward (base → trained)")
L.append("")
L.append("| Component | Weight | Base | Trained | Δ |")
L.append("|---|---|---|---|---|")
for key, short, w in COMPONENTS:
L.append(f"| {short} | {w:.2f} | {_fmt(bc[key])} | {_fmt(tc[key])} | "
f"**{_delta_str(bc[key], tc[key])}** |")
L.append("")
L.append("> `reversibility_correct` is the headline skill (weight 0.30). A positive Δ "
"here is the most direct evidence the model learned to predict reversibility.")
L.append("")
# 2. Per-tier
L.append("## 2. Per-tier reward + reversibility_correct")
L.append("")
L.append("| Tier | n | Base reward | Trained reward | Δ reward | Base rev_correct | Trained rev_correct | Δ rev_correct |")
L.append("|---|---|---|---|---|---|---|---|")
for tier in sorted(set(bt) | set(tt)):
b = bt.get(tier, {})
t = tt.get(tier, {})
L.append(
f"| T{tier} | {b.get('n', t.get('n', '?'))} | "
f"{_fmt(b.get('reward'))} | {_fmt(t.get('reward'))} | "
f"**{_delta_str(b.get('reward'), t.get('reward'))}** | "
f"{_fmt(b.get('reversibility_correct'))} | {_fmt(t.get('reversibility_correct'))} | "
f"**{_delta_str(b.get('reversibility_correct'), t.get('reversibility_correct'))}** |"
)
L.append("")
# 3. Action-level reversibility prediction accuracy
L.append("## 3. Reversibility PREDICTION accuracy (action-level)")
L.append("")
L.append("Every action carries `predicted_reversibility` vs `ground_truth_reversibility`. "
"This is the direct measure of whether the model can *classify* reversibility — "
"split by ground-truth class because getting IRREVERSIBLE right is safety-critical.")
L.append("")
L.append("| Subset | Base acc | Trained acc | Δ | n (base/trained) |")
L.append("|---|---|---|---|---|")
L.append(f"| Overall | {_fmt(ba['overall_acc'])} | {_fmt(ta['overall_acc'])} | "
f"**{_delta_str(ba['overall_acc'], ta['overall_acc'])}** | {ba['overall_n']}/{ta['overall_n']} |")
L.append(f"| On irreversible ops | {_fmt(ba['irreversible_acc'])} | {_fmt(ta['irreversible_acc'])} | "
f"**{_delta_str(ba['irreversible_acc'], ta['irreversible_acc'])}** | {ba['irreversible_n']}/{ta['irreversible_n']} |")
L.append(f"| On reversible ops | {_fmt(ba['reversible_acc'])} | {_fmt(ta['reversible_acc'])} | "
f"**{_delta_str(ba['reversible_acc'], ta['reversible_acc'])}** | {ba['reversible_n']}/{ta['reversible_n']} |")
L.append("")
# 4. Confidence calibration
L.append("## 4. Confidence calibration (mean stated confidence)")
L.append("")
L.append("| | Base | Trained | Δ |")
L.append("|---|---|---|---|")
L.append(f"| Confidence when prediction CORRECT | {_fmt(ba['conf_when_right'])} | {_fmt(ta['conf_when_right'])} | **{_delta_str(ba['conf_when_right'], ta['conf_when_right'])}** |")
L.append(f"| Confidence when prediction WRONG | {_fmt(ba['conf_when_wrong'])} | {_fmt(ta['conf_when_wrong'])} | **{_delta_str(ba['conf_when_wrong'], ta['conf_when_wrong'])}** |")
L.append("")
L.append("> Healthy calibration: high confidence when right, LOWER confidence when wrong. "
"If training reduced 'confidence when wrong', the model became better at knowing "
"when it doesn't know.")
L.append("")
# 5. Action-type distribution
L.append("## 5. Action-type distribution (all steps)")
L.append("")
L.append("| Action | Base | Trained | Δ |")
L.append("|---|---|---|---|")
for at in sorted(set(bact) | set(tact)):
b = bact.get(at, 0)
t = tact.get(at, 0)
L.append(f"| {at} | {b} | {t} | **{t - b:+d}** |")
L.append("")
return "\n".join(L)
def main() -> None:
p = argparse.ArgumentParser()
p.add_argument("--prefix", required=True, help="Model log prefix, e.g. qwen7b or llama8b")
p.add_argument("--results-dir", default="eval/results", help="Directory holding the inference JSONs")
args = p.parse_args()
results_dir = Path(args.results_dir)
report = build_report(args.prefix, results_dir)
out_path = results_dir / f"{args.prefix}_capability_report.md"
out_path.write_text(report + "\n")
print(report)
print(f"\n[written] {out_path}")
if __name__ == "__main__":
main()
|