StepProbe / scripts /compute_ci.py
Akiyue's picture
Add files using upload-large-folder tool
3ccaf5a verified
Raw
History Blame Contribute Delete
10.3 kB
"""Bootstrap confidence intervals and paired significance tests.
Reads the per-problem diagnosed traces in results/diagnosis/ and, for each
(model, benchmark, quantization) cell, computes:
- 95% bootstrap CIs for accuracy, average FFS, and ECR (5000 resamples).
- A full SSR curve CI band (per-depth bootstrap).
- For (base, restored) pairs: a paired bootstrap significance test on
ΔAccuracy (uses the SAME sampled problem ids for base and restored in
each resample, so the null cancels properly).
Emits one *_ci.json next to each existing *_metrics.json, plus one pair-level
*_sig.json per (model, benchmark, method) restoration comparison.
These are real, defensible uncertainty estimates given a SINGLE inference run:
they capture variance across problems, not variance across seeds. For a Q1
paper you want both — this script delivers the first half for free; the
seed-variance half requires re-running inference with NUM_RUNS>=3.
"""
from __future__ import annotations
import argparse
import glob
import json
import os
import re
from typing import Dict, List, Optional, Tuple
import numpy as np
BENCHMARKS = ("gsm8k", "math500", "gpqa")
QUANT_METHODS = ("awq_w4", "gptq_w4", "bnb_nf4_w4")
# ---------------------------------------------------------------------------
# Per-problem metric extraction
# ---------------------------------------------------------------------------
def first_failure_step(steps: List[dict]) -> Optional[int]:
for s in steps:
if s.get("is_correct") is False:
return s.get("index")
return None
def cascade_rate(steps: List[dict]) -> Optional[float]:
"""Fraction of steps AFTER the first failure that are also incorrect.
Returns None for problems with no failure (so the aggregate only averages
over problems that actually failed — matches aggregate_metrics.ecr)."""
ffs = first_failure_step(steps)
if ffs is None:
return None
tail = [s for s in steps if s.get("index", 0) > ffs]
if not tail:
return None
return sum(1 for s in tail if s.get("is_correct") is False) / len(tail)
def survival_at_depth(steps: List[dict], depth: int) -> Optional[bool]:
"""True if all of steps[0..depth] are correct, None if the trace is shorter."""
if len(steps) <= depth:
return None
return all(s.get("is_correct", True) for s in steps[: depth + 1])
def load_per_problem(jsonl_path: str) -> List[dict]:
"""Return a list of per-problem records with the fields bootstrapping needs."""
out = []
with open(jsonl_path) as f:
for line in f:
t = json.loads(line)
steps = t.get("steps", []) or []
out.append({
"problem_id": t.get("problem_id"),
"is_correct": bool(t.get("is_correct_final", False)),
"ffs": first_failure_step(steps),
"ecr": cascade_rate(steps),
"step_correct": [s.get("is_correct", True) for s in steps],
})
return out
# ---------------------------------------------------------------------------
# Bootstrap utilities
# ---------------------------------------------------------------------------
def _ci(samples: np.ndarray, alpha: float = 0.05) -> Tuple[float, float]:
lo, hi = np.percentile(samples, [100 * alpha / 2, 100 * (1 - alpha / 2)])
return float(lo), float(hi)
def bootstrap_single(records: List[dict], n_boot: int = 5000,
rng: Optional[np.random.Generator] = None) -> dict:
"""Bootstrap accuracy, avg_ffs, ecr, and the SSR curve for one cell."""
rng = rng or np.random.default_rng(0)
n = len(records)
if n == 0:
return {}
acc_vec = np.array([r["is_correct"] for r in records], dtype=float)
ffs_vec = np.array([r["ffs"] if r["ffs"] is not None else np.nan for r in records], dtype=float)
ecr_vec = np.array([r["ecr"] if r["ecr"] is not None else np.nan for r in records], dtype=float)
# SSR per depth: for problems with enough steps, is the trace "alive" at depth d?
max_depth = 30
alive = np.full((n, max_depth), np.nan)
for i, r in enumerate(records):
steps = r["step_correct"]
for d in range(max_depth):
if d < len(steps):
alive[i, d] = float(all(steps[: d + 1]))
acc_samples = np.empty(n_boot)
ffs_samples = np.empty(n_boot)
ecr_samples = np.empty(n_boot)
ssr_samples = np.full((n_boot, max_depth), np.nan)
for b in range(n_boot):
idx = rng.integers(0, n, size=n)
acc_samples[b] = acc_vec[idx].mean()
f = ffs_vec[idx]
ffs_samples[b] = np.nanmean(f) if np.any(~np.isnan(f)) else np.nan
e = ecr_vec[idx]
ecr_samples[b] = np.nanmean(e) if np.any(~np.isnan(e)) else np.nan
for d in range(max_depth):
col = alive[idx, d]
valid = ~np.isnan(col)
if valid.any():
ssr_samples[b, d] = col[valid].mean()
def summarize(x):
x = x[~np.isnan(x)]
if x.size == 0:
return None
lo, hi = _ci(x)
return {"mean": float(x.mean()), "ci_lo": lo, "ci_hi": hi}
ssr_ci_lo = np.nanpercentile(ssr_samples, 2.5, axis=0)
ssr_ci_hi = np.nanpercentile(ssr_samples, 97.5, axis=0)
ssr_mean = np.nanmean(ssr_samples, axis=0)
return {
"n": int(n),
"accuracy": summarize(acc_samples),
"avg_ffs": summarize(ffs_samples),
"ecr": summarize(ecr_samples),
"ssr_curve_ci": {
"mean": [None if np.isnan(v) else float(v) for v in ssr_mean],
"ci_lo": [None if np.isnan(v) else float(v) for v in ssr_ci_lo],
"ci_hi": [None if np.isnan(v) else float(v) for v in ssr_ci_hi],
},
}
def paired_bootstrap_sig(base: List[dict], rest: List[dict],
n_boot: int = 5000,
rng: Optional[np.random.Generator] = None) -> dict:
"""Paired bootstrap on ΔAccuracy between base and restored.
Problems are paired by problem_id. For each bootstrap resample we draw the
SAME indices in both conditions, compute the delta, and accumulate. The
two-sided p-value is 2 * min(P(Δ<=0), P(Δ>=0)).
"""
rng = rng or np.random.default_rng(0)
base_by_id = {r["problem_id"]: r for r in base}
rest_by_id = {r["problem_id"]: r for r in rest}
shared = sorted(set(base_by_id) & set(rest_by_id))
if not shared:
return {"n_pairs": 0}
b_vec = np.array([base_by_id[pid]["is_correct"] for pid in shared], dtype=float)
r_vec = np.array([rest_by_id[pid]["is_correct"] for pid in shared], dtype=float)
n = len(shared)
deltas = np.empty(n_boot)
for i in range(n_boot):
idx = rng.integers(0, n, size=n)
deltas[i] = r_vec[idx].mean() - b_vec[idx].mean()
observed = float(r_vec.mean() - b_vec.mean())
p_two = 2 * min((deltas <= 0).mean(), (deltas >= 0).mean())
lo, hi = _ci(deltas)
return {
"n_pairs": int(n),
"delta_acc_observed": observed,
"delta_acc_ci_lo": lo,
"delta_acc_ci_hi": hi,
"p_value": float(p_two),
}
# ---------------------------------------------------------------------------
# CLI glue
# ---------------------------------------------------------------------------
def _discover_cells(diagnosis_dir: str):
"""Yield (model, benchmark, quant, jsonl_path) tuples for everything on disk."""
for quant_dir in sorted(glob.glob(os.path.join(diagnosis_dir, "*"))):
if not os.path.isdir(quant_dir):
continue
quant = os.path.basename(quant_dir)
for model_dir in sorted(glob.glob(os.path.join(quant_dir, "*"))):
if not os.path.isdir(model_dir):
continue
model = os.path.basename(model_dir)
for jsonl in sorted(glob.glob(os.path.join(model_dir, "*_run0.jsonl"))):
base = os.path.basename(jsonl)
m = re.match(r"(.+)_run0\.jsonl$", base)
if not m:
continue
bench = m.group(1)
if bench not in BENCHMARKS:
continue
yield model, bench, quant, jsonl
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--diagnosis", default="results/diagnosis",
help="Root diagnosis dir (quant/model/<bench>_run0.jsonl)")
parser.add_argument("--output", default="results/metrics",
help="Where to drop *_ci.json and *_sig.json")
parser.add_argument("--n-boot", type=int, default=5000)
parser.add_argument("--seed", type=int, default=0)
args = parser.parse_args()
os.makedirs(args.output, exist_ok=True)
rng = np.random.default_rng(args.seed)
# --- Phase A: single-cell CIs ------------------------------------------
cells = list(_discover_cells(args.diagnosis))
print(f"Found {len(cells)} diagnosis cells to bootstrap")
per_problem_cache: Dict[Tuple[str, str, str], List[dict]] = {}
for model, bench, quant, jsonl in cells:
records = load_per_problem(jsonl)
per_problem_cache[(model, bench, quant)] = records
ci = bootstrap_single(records, n_boot=args.n_boot, rng=rng)
out_path = os.path.join(args.output, f"{model}_{quant}_{bench}_run0_ci.json")
with open(out_path, "w") as f:
json.dump(ci, f, indent=2)
print(f" wrote {len(cells)} *_ci.json files")
# --- Phase B: paired significance (base vs restored) --------------------
n_sig = 0
for (model, bench, quant), base_recs in per_problem_cache.items():
if quant.endswith("_restored"):
continue
rest_recs = per_problem_cache.get((model, bench, quant + "_restored"))
if not rest_recs:
continue
sig = paired_bootstrap_sig(base_recs, rest_recs, n_boot=args.n_boot, rng=rng)
out_path = os.path.join(args.output, f"{model}_{quant}_{bench}_run0_sig.json")
with open(out_path, "w") as f:
json.dump(sig, f, indent=2)
n_sig += 1
print(f" wrote {n_sig} *_sig.json files (paired base vs restored)")
if __name__ == "__main__":
main()