Download code/make_spec.py from AmareshHebbar/Modeldiffr_Bench: direct link, hf CLI and curl.
- Browser
- Download file 3.48 kB
-
https://huggingface.co/datasets/AmareshHebbar/Modeldiffr_Bench/resolve/main/code/make_spec.py
- Command line
-
hf download hf://datasets/AmareshHebbar/Modeldiffr_Bench/code/make_spec.py
-
curl -L -o make_spec.py https://huggingface.co/datasets/AmareshHebbar/Modeldiffr_Bench/resolve/main/code/make_spec.py
3.48 kB
| import argparse | |
| import os | |
| from pathlib import Path | |
| import yaml | |
| HERE = Path(__file__).resolve().parent | |
| CONTROL_BASE = "Qwen/Qwen2.5-0.5B-Instruct" | |
| REAL = [ | |
| ("shortthink-sft", "sft", "Qwen/Qwen3-1.7B", "AmareshHebbar/ShortThink_Qwen3_1.7B_SFT"), | |
| ("shortthink-grpo", "grpo", "Qwen/Qwen3-1.7B", "AmareshHebbar/ShortThink_Qwen3_1.7B_GRPO"), | |
| ("toolsure-orpo", "orpo", "Qwen/Qwen2.5-1.5B-Instruct", "AmareshHebbar/ToolSure_Qwen2.5_1.5B_ORPO"), | |
| ("stretch32k-sft", "long context sft", "TinyLlama/TinyLlama_v1.1", "AmareshHebbar/Stretch32K_TinyLlama_1.1B_SFT"), | |
| ("instruct-qwen2.5-0.5b", "instruction tuning", "Qwen/Qwen2.5-0.5B", "Qwen/Qwen2.5-0.5B-Instruct"), | |
| ] | |
| BITNARROW = [ | |
| ("bitnarrow-qwen7b", "weight edit", "Qwen/Qwen2.5-7B-Instruct", "qwen7b_narrow"), | |
| ("bitnarrow-llama8b", "weight edit", "meta-llama/Llama-3.1-8B-Instruct", "llama8b_narrow"), | |
| ] | |
| def controls() -> list[dict]: | |
| out = [] | |
| for s in range(3): | |
| out.append({"name": f"null-identity-s{s}", "kind": "null identity", "control": {"kind": "identity", "seed": s}, "seed": s}) | |
| for s in range(3): | |
| out.append({ | |
| "name": f"null-micro-noise-s{s}", | |
| "kind": "null micro noise", | |
| "control": {"kind": "noise", "strength": 0.0001, "seed": s}, | |
| "expect": "no_change", | |
| "seed": s, | |
| }) | |
| planted = [ | |
| ("noise-0.10", {"kind": "noise", "strength": 0.10}), | |
| ("noise-0.20", {"kind": "noise", "strength": 0.20}), | |
| ("noise-0.40", {"kind": "noise", "strength": 0.40}), | |
| ("prune-0.30", {"kind": "prune", "strength": 0.30}), | |
| ("zero-last-mlp", {"kind": "zero", "match": r"layers\.23\.mlp\.down_proj\.weight$"}), | |
| ("scale-mlp-0.5", {"kind": "scale", "strength": 0.5, "match": r"mlp\.down_proj\.weight$"}), | |
| ] | |
| for name, c in planted: | |
| out.append({"name": f"planted-{name}", "kind": "planted " + c["kind"], "control": c}) | |
| for e in out: | |
| e["base"] = CONTROL_BASE | |
| return out | |
| def main() -> None: | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--out", default=str(HERE / "spec.yaml")) | |
| ap.add_argument("--exports", default=os.environ.get("BITNARROW_EXPORTS", "")) | |
| ap.add_argument("--no-real", action="store_true") | |
| ap.add_argument("--smoke", action="store_true") | |
| ap.add_argument("--dtype", default=os.environ.get("MD_DTYPE", "bfloat16")) | |
| args = ap.parse_args() | |
| audits = controls() | |
| if args.smoke: | |
| audits = [a for a in audits if a["name"] in ("null-micro-noise-s0", "planted-noise-0.40")] | |
| if not args.no_real and not args.smoke: | |
| for name, kind, base, deriv in REAL: | |
| audits.append({"name": f"real-{name}", "kind": kind, "base": base, "derivative": deriv, "expect": "change"}) | |
| if args.exports: | |
| for name, kind, base, folder in BITNARROW: | |
| path = Path(args.exports) / folder | |
| if path.exists(): | |
| audits.append({"name": f"real-{name}", "kind": kind, "base": base, "derivative": str(path), "expect": "change"}) | |
| spec = { | |
| "out": "bench/results/audits", | |
| "defaults": {"suite": "lite", "dtype": args.dtype, "seed": 0, "limit": 8 if args.smoke else None}, | |
| "audits": audits, | |
| } | |
| if spec["defaults"]["limit"] is None: | |
| del spec["defaults"]["limit"] | |
| Path(args.out).write_text(yaml.safe_dump(spec, sort_keys=False), encoding="utf-8") | |
| print(f"{args.out}: {len(audits)} audits") | |
| if __name__ == "__main__": | |
| main() | |