File size: 1,597 Bytes
ccbd209
 
 
 
 
 
 
 
 
9bf4115
 
 
ccbd209
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Critique -> revise: the model learns from criticism.

For each prompt: the policy answers, the judge writes a criticism, the policy REVISES
using that criticism, and if the revision scores higher it becomes a training pair
(prompt -> revised). This is how "take criticism into consideration" becomes signal.
"""


def _gen(model, tok, messages, max_new_tokens=512):
    from core.genutil import chat_generate
    return chat_generate(model, tok, messages, max_new_tokens=max_new_tokens,
                         do_sample=True, temperature=0.7, top_p=0.95)


def critique_revise_batch(policy, tok, judge, prompts):
    """Returns improved (instruction, response) pairs where revision beat the original."""
    pairs = []
    for p in prompts:
        first = _gen(policy, tok, [{"role": "user", "content": p}])
        verdict = judge.critique(p, first)
        if verdict["score"] >= 0.9:
            pairs.append({"instruction": p, "response": first, "score": verdict["score"]})
            continue
        revise_msgs = [
            {"role": "user", "content": p},
            {"role": "assistant", "content": first},
            {"role": "user", "content": f"Criticism: {verdict['criticism']} "
                                         f"(weakest: {verdict['weakest']}). Revise to fix this."},
        ]
        revised = _gen(policy, tok, revise_msgs)
        rscore = judge.score(p, revised)
        best, bscore = (revised, rscore) if rscore > verdict["score"] else (first, verdict["score"])
        pairs.append({"instruction": p, "response": best, "score": bscore})
    return pairs