| """Post-training stage probe (reproduces Section 7.2 / Appendix P).
|
|
|
| Loads the GPT-4.1 and Kimi-K2 judge CSVs in
|
| ``probe1_post_training/judges/``, averages the two judges' DS>=2 binary
|
| outcomes per response, and prints per-checkpoint per-family rates plus
|
| a Trap column (mean of F1+F2). Checkpoints span base / SFT / DPO / RLHF
|
| variants for OLMo, Qwen, and Mistral families.
|
| """
|
|
|
| from __future__ import annotations
|
|
|
| import glob
|
| import re
|
| import sys
|
| from pathlib import Path
|
|
|
| import pandas as pd
|
|
|
| sys.path.insert(0, str(Path(__file__).resolve().parent))
|
| from _paths import PROBE1
|
|
|
| CHECKPOINT_LABEL = {
|
| "probe1_local_olmo2-7b-base": "olmo2-7b-base",
|
| "probe1_local_olmo2-7b-sft": "olmo2-7b-sft",
|
| "probe1_local_olmo2-7b-dpo": "olmo2-7b-dpo",
|
| "probe1_local_qwen25-7b-base": "qwen25-7b-base",
|
| "probe1_qwen25_7b_instruct": "qwen25-7b-instruct",
|
| "probe1_mistral7b_v02_instruct": "mistral7b-v02-instruct",
|
| "probe1_mistral7b_v03_instruct": "mistral7b-v03-instruct",
|
| }
|
|
|
|
|
| ORDER = [
|
| "olmo2-7b-base", "olmo2-7b-sft", "olmo2-7b-dpo",
|
| "qwen25-7b-base", "qwen25-7b-instruct",
|
| "mistral7b-v02-instruct", "mistral7b-v03-instruct",
|
| ]
|
|
|
|
|
| def _family_from_id(scenario_id: str) -> int:
|
| m = re.match(r"^F(\d)-", scenario_id)
|
| return int(m.group(1)) if m else -1
|
|
|
|
|
| def main() -> None:
|
| files = sorted(glob.glob(str(PROBE1 / "judges" / "*.csv")))
|
| if not files:
|
| print("No probe1 judge CSVs found.")
|
| return
|
|
|
| frames = []
|
| for f in files:
|
| df = pd.read_csv(f)
|
| frames.append(df)
|
| cat = pd.concat(frames, ignore_index=True)
|
|
|
| cat["fam"] = cat["scenario_id"].apply(_family_from_id)
|
|
|
|
|
|
|
|
|
| per_resp = (cat.groupby(["response_type", "scenario_id", "fam"])
|
| ["displacement_severity"].mean().reset_index())
|
| per_resp["disp_bin"] = (per_resp["displacement_severity"] >= 2.0).astype(int)
|
|
|
| name_w = max(len(v) for v in ORDER)
|
| header = (f"{'Checkpoint'.ljust(name_w)} "
|
| f"{'F1':>6} {'F2':>6} {'F3':>6} {'F0':>6} {'Trap':>6}")
|
| print(header)
|
| print("-" * len(header))
|
|
|
| for ckpt in ORDER:
|
| rt = next((k for k, v in CHECKPOINT_LABEL.items() if v == ckpt), None)
|
| if rt is None:
|
| continue
|
| sub = per_resp[per_resp["response_type"] == rt]
|
| if len(sub) == 0:
|
| continue
|
| rates = {}
|
| for fam in [1, 2, 3, 0]:
|
| cell = sub[sub["fam"] == fam]
|
| rates[fam] = cell["disp_bin"].mean() if len(cell) else float("nan")
|
| trap = (rates[1] + rates[2]) / 2.0
|
| print(f"{ckpt.ljust(name_w)} "
|
| f"{rates[1] * 100:5.1f}% {rates[2] * 100:5.1f}% "
|
| f"{rates[3] * 100:5.1f}% {rates[0] * 100:5.1f}% "
|
| f"{trap * 100:5.1f}%")
|
|
|
|
|
| if __name__ == "__main__":
|
| main()
|
|
|
