src/eval_prob.py

"""Fast per-question evaluator for H2 revival curves.

For every forget10 question: teacher-forced answer probability
(exp(-avg_token_loss), same construction as OpenUnlearning's
forget_Q_A_Prob — parity-checked against their TOFU_EVAL.json), plus the
free-energy signal logsumexp(logits) at the last prompt position.
Optional --rouge: greedy generation (max_new_tokens=200) + ROUGE-L recall.

Output json: {"prob": {idx: p}, "energy": {idx: e}, "rouge": {idx: r}?}

Usage: python eval_prob.py --model_dir M --out out.json [--rouge] [--indices 0,1,...]
"""

import argparse
import json

import torch
import transformers

SYSTEM_PROMPT = "You are a helpful assistant."
DATE_STRING = "10 Apr 2025"


def build_rows():
    from datasets import load_dataset

    ds = load_dataset("locuslab/TOFU", "forget10")["train"]
    return [(i, ds[i]["question"], ds[i]["answer"]) for i in range(len(ds))]


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model_dir", required=True)
    ap.add_argument("--out", required=True)
    ap.add_argument("--rouge", action="store_true")
    ap.add_argument("--indices", default=None, help="comma list; default all 400")
    ap.add_argument("--batch", type=int, default=16)
    args = ap.parse_args()

    tok = transformers.AutoTokenizer.from_pretrained(args.model_dir)
    tok.padding_side = "right"
    if tok.pad_token is None:
        tok.pad_token = tok.eos_token
    hf = transformers.AutoModelForCausalLM.from_pretrained(
        args.model_dir, dtype=torch.bfloat16
    ).cuda()
    hf.eval()

    rows = build_rows()
    if args.indices:
        keep = {int(x) for x in args.indices.split(",")}
        rows = [r for r in rows if r[0] in keep]

    def encode(question, answer):
        chat_sys = [{"role": "system", "content": SYSTEM_PROMPT},
                    {"role": "user", "content": question}]
        prompt_ids = tok.apply_chat_template(
            chat_sys, tokenize=True, add_generation_prompt=True,
            date_string=DATE_STRING)["input_ids"]
        full_ids = tok.apply_chat_template(
            chat_sys + [{"role": "assistant", "content": answer}],
            tokenize=True, add_generation_prompt=False,
            date_string=DATE_STRING)["input_ids"]
        return prompt_ids, full_ids

    prob, energy = {}, {}
    with torch.no_grad():
        for b0 in range(0, len(rows), args.batch):
            chunk = rows[b0 : b0 + args.batch]
            enc = [encode(q, a) for _, q, a in chunk]
            maxlen = max(len(f) for _, f in enc)
            ids = torch.full((len(chunk), maxlen), tok.pad_token_id, dtype=torch.long)
            for j, (_, f) in enumerate(enc):
                ids[j, : len(f)] = torch.tensor(f)
            ids = ids.cuda()
            attn = (ids != tok.pad_token_id).long()
            logits = hf(ids, attention_mask=attn).logits.float()
            logprobs = torch.log_softmax(logits, dim=-1)
            for j, ((idx, _, _), (p_ids, f_ids)) in enumerate(zip(chunk, enc)):
                n_p, n_f = len(p_ids), len(f_ids)
                # answer tokens are positions n_p..n_f-1, predicted from n_p-1..n_f-2
                tgt = torch.tensor(f_ids[n_p:], device="cuda")
                lp = logprobs[j, n_p - 1 : n_f - 1].gather(1, tgt[:, None]).squeeze(1)
                prob[idx] = float(torch.exp(lp.mean()))
                energy[idx] = float(torch.logsumexp(logits[j, n_p - 1], dim=-1))

    out = {"prob": prob, "energy": energy}

    if args.rouge:
        from rouge_score import rouge_scorer

        scorer = rouge_scorer.RougeScorer(["rougeL"], use_stemmer=True)
        tok.padding_side = "left"
        rouge = {}
        with torch.no_grad():
            for b0 in range(0, len(rows), args.batch):
                chunk = rows[b0 : b0 + args.batch]
                prompts = [
                    tok.apply_chat_template(
                        [{"role": "system", "content": SYSTEM_PROMPT},
                         {"role": "user", "content": q}],
                        tokenize=False, add_generation_prompt=True,
                        date_string=DATE_STRING)
                    for _, q, _ in chunk
                ]
                enc = tok(prompts, return_tensors="pt", padding=True,
                          add_special_tokens=False).to("cuda")
                gen = hf.generate(**enc, max_new_tokens=200, do_sample=False,
                                  pad_token_id=tok.pad_token_id)
                for j, (idx, _, a) in enumerate(chunk):
                    text = tok.decode(gen[j, enc["input_ids"].shape[1]:],
                                      skip_special_tokens=True)
                    rouge[idx] = scorer.score(a, text)["rougeL"].recall
        out["rouge"] = rouge

    json.dump(out, open(args.out, "w"))
    print("saved", args.out, f"n={len(prob)}")


if __name__ == "__main__":
    main()