XINLI1997 commited on
Commit
7f924e0
·
verified ·
1 Parent(s): 8b292b7

Model card: locked protocol grades every response with the GPT-4.1-mini judge

Browse files
Files changed (1) hide show
  1. README.md +1 -1
README.md CHANGED
@@ -34,7 +34,7 @@ GRPO-trained reference checkpoint from **WirelessMathBench-XL: An Auditable Benc
34
  | Full 800-item test split (paper Tab. 3, row D10) | 47.88% |
35
  | 310-item public test subset (paper Appendix L) | 46.45% |
36
 
37
- Locked protocol: 2k-token answer budget, T = 0.6, hierarchical verifier (deterministic match, canonicalisation, GPT-4.1-mini fallback judge).
38
 
39
  Training-time check under greedy model selection (paper Tab. 7): base 21.88% → GRPO 39.50%. These values use a different protocol from the locked scores.
40
 
 
34
  | Full 800-item test split (paper Tab. 3, row D10) | 47.88% |
35
  | 310-item public test subset (paper Appendix L) | 46.45% |
36
 
37
+ Locked protocol: 2k-token answer budget, T = 0.6, GPT-4.1-mini judge on every response.
38
 
39
  Training-time check under greedy model selection (paper Tab. 7): base 21.88% → GRPO 39.50%. These values use a different protocol from the locked scores.
40