File size: 3,136 Bytes
ccbd209
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9bf4115
ccbd209
 
9bf4115
ccbd209
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9bd725a
 
 
 
ccbd209
9bd725a
ccbd209
9bd725a
ccbd209
9bd725a
 
 
 
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
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
68
69
70
71
72
73
74
75
76
77
"""The rater/critic: scores a response 0..1 against the rubric and writes criticism.

This is the reward signal for GRPO AND the criticism source for critique-revise.
Backend-flexible: loads JUDGE_MODEL via transformers. Swap to an API if you prefer
a stronger grader -- just keep score()/critique() returning the same shape.
"""
import json
import os
import re
from pathlib import Path

ROOT = Path(__file__).resolve().parent.parent


def load_rubric():
    rubric = (ROOT / "rubric.md").read_text(encoding="utf-8")
    fb = ROOT / "feedback" / "criticisms.jsonl"
    if fb.exists():
        crits = [json.loads(l)["criticism"] for l in fb.read_text(encoding="utf-8").splitlines() if l.strip()]
        if crits:
            rubric += "\n\n## Accumulated user/judge criticisms (avoid these)\n" + \
                      "\n".join(f"- {c}" for c in crits[-50:])
    return rubric


_SYS = ("You are a strict, fair grader. Score the RESPONSE to the PROMPT against the rubric. "
        "Return ONLY JSON: {\"score\": float 0..1, \"weakest\": str, \"criticism\": str}.")


class Judge:
    def __init__(self, model_name=None, device="cuda"):
        from transformers import AutoModelForCausalLM, AutoTokenizer
        import torch
        from config import JUDGE_MODEL
        self.rubric = load_rubric()
        self.tok = AutoTokenizer.from_pretrained(model_name or JUDGE_MODEL)
        self.model = AutoModelForCausalLM.from_pretrained(
            model_name or JUDGE_MODEL, torch_dtype=torch.bfloat16, device_map=device)

    def _ask(self, prompt, response):
        from core.genutil import chat_generate
        msg = [{"role": "system", "content": _SYS},
               {"role": "user", "content": f"RUBRIC:\n{self.rubric}\n\nPROMPT:\n{prompt}\n\nRESPONSE:\n{response}"}]
        text = chat_generate(self.model, self.tok, msg, max_new_tokens=200, do_sample=False)
        m = re.search(r"\{.*\}", text, re.DOTALL)
        try:
            d = json.loads(m.group(0)) if m else {}
        except Exception:
            d = {}
        return {"score": float(max(0.0, min(1.0, d.get("score", 0.0)))),
                "weakest": d.get("weakest", "unknown"),
                "criticism": d.get("criticism", "")}

    def score(self, prompt, response):
        return self._ask(prompt, response)["score"]

    def critique(self, prompt, response):
        return self._ask(prompt, response)


def make_reward_func(judge):
    """TRL GRPO reward_func: (prompts, completions, **kw) -> list[float] in 0..1.
    Routes each item through its modality verifier (objective) blended with the judge."""
    from core import modalities

    def reward(prompts, completions, **kw):
        types = kw.get("type") or [None] * len(prompts)
        out = []
        for i, (p, c) in enumerate(zip(prompts, completions)):
            text = c if isinstance(c, str) else c[-1]["content"]
            ptext = p if isinstance(p, str) else p[-1]["content"]
            js = judge.score(ptext, text)
            t = types[i] if i < len(types) else None
            out.append(modalities.blended_reward(t, js, text))
        return out
    return reward