File size: 3,161 Bytes
fd68fe9 | 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 | # /// script
# dependencies = ["accelerate>=1,<2", "datasets>=3,<5", "huggingface-hub>=0.26,<2", "peft>=0.13,<1", "torch>=2.4", "transformers>=4.46,<5"]
# ///
"""Evaluate a base model or LoRA candidate on a frozen curated split."""
from __future__ import annotations
from datetime import UTC, datetime
import hashlib
import json
import os
import re
from datasets import load_dataset
from huggingface_hub import HfApi
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
def _answer(text: str) -> str:
matches = re.findall(r"[-+]?\d+(?:\.\d+)?(?:/\d+)?", text.replace(",", ""))
return matches[-1] if matches else text.strip().casefold()[-200:]
def main() -> None:
model_repo = os.environ["MODEL_REPO"]
base_repo = os.getenv("BASE_MODEL_REPO")
revision = os.getenv("MODEL_REVISION") or None
dataset_repo = os.environ["DATASET_REPO"]
split = os.getenv("DATASET_SPLIT", "test")
max_cases = int(os.getenv("MAX_CASES", "500"))
tokenizer = AutoTokenizer.from_pretrained(base_repo or model_repo, revision=revision, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(base_repo or model_repo, revision=revision, torch_dtype="auto", device_map="auto")
if base_repo:
from peft import PeftModel
model = PeftModel.from_pretrained(model, model_repo)
data = load_dataset(dataset_repo, split=split, streaming=True)
cases, passed, unsafe = [], 0, 0
for row in data.take(max_cases):
prompt = f"Solve the task. Return a clear final answer.\n\n{row['input']}"
encoded = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096).to(model.device)
with torch.inference_mode():
output = model.generate(**encoded, max_new_tokens=1024, do_sample=False)
text = tokenizer.decode(output[0][encoded["input_ids"].shape[1]:], skip_special_tokens=True)
reference = str((row.get("evaluation") or {}).get("reference_answer") or "")
ok = bool(reference) and _answer(reference) == _answer(text)
is_unsafe = any(token in text for token in ("hf_", "sk-"))
passed += ok
unsafe += is_unsafe
cases.append({"example_id": row["example_id"], "passed": ok, "unsafe": is_unsafe, "output_sha256": hashlib.sha256(text.encode()).hexdigest()})
report = {
"schema_version": "1.0.0", "created_at": datetime.now(UTC).isoformat(),
"model": {"repo_id": model_repo, "base_repo_id": base_repo, "revision": revision},
"dataset": {"repo_id": dataset_repo, "split": split},
"metrics": {"case_count": len(cases), "pass_rate": passed / len(cases) if cases else 0.0, "unsafe_rate": unsafe / len(cases) if cases else 0.0},
"cases": cases,
}
payload = json.dumps(report, sort_keys=True, indent=2) + "\n"
HfApi().create_repo(os.environ["REPORT_REPO"], repo_type="dataset", private=True, exist_ok=True)
HfApi().upload_file(path_or_fileobj=payload.encode(), path_in_repo=os.environ["REPORT_PATH"], repo_id=os.environ["REPORT_REPO"], repo_type="dataset", commit_message=f"Evaluate {model_repo}")
print(payload)
if __name__ == "__main__":
main()
|