File size: 1,467 Bytes
ccbd209
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9bf4115
9bd725a
ccbd209
 
9bf4115
 
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
"""Collapse guard: the held-out, human-grounded eval gate.

The whole self-refinement loop is dangerous precisely because the AI rates its own
training data -- that's how model collapse happens. This gate is the antidote: a FROZEN
eval set that the AI never generates or rates, scored by the judge, used only to decide
whether to KEEP a refinement round. If a round regresses on real held-out tasks, it's
reverted. This converts "spirals down" into "only changes that actually help survive."
"""
import json
from pathlib import Path


def load_eval(path):
    p = Path(path)
    if not p.exists():
        return []
    return [json.loads(l) for l in p.read_text(encoding="utf-8").splitlines() if l.strip()]


def evaluate(policy, tok, judge, eval_items):
    """Mean judge score of the policy on the frozen eval set."""
    if not eval_items:
        return 0.0
    from core.genutil import chat_generate
    from core import modalities
    total = 0.0
    for it in eval_items:
        resp = chat_generate(policy, tok, [{"role": "user", "content": it["instruction"]}],
                             max_new_tokens=512, do_sample=False)
        js = judge.score(it["instruction"], resp)
        total += modalities.blended_reward(it.get("type"), js, resp)
    return total / len(eval_items)


def keep_round(new_score, prev_score, tolerance=0.0):
    """Keep the round only if it doesn't regress beyond tolerance."""
    return new_score >= (prev_score - tolerance)