Training setup
- Prompts:
U-WIN/equational-theory-lean2nl-references-gpt-approved-v2-rubric @ fd284c793204 (2182 problems, all TRUE; 54 finite-carrier).
- Eval:
U-WIN/sair-stage2-nl-verified-v2-rubric @ ed7fb4630dfe (205 problems, 99 TRUE / 106 FALSE), 4 samples per problem every 20 rollouts.
- Reward: rubric grading by a frozen local Qwen3.8-27B judge (policy id
v2rubric-verdict-neutral-user-rubric-20261104-v3): generic policy + two-field JSON output (assessment, score_0_to_7), reward = score/7, dedicated verdict lines removed from the candidate, no reference solution or truth label shown to the judge.
- Sampling: temperature 1.0, 32,768 response tokens, GRPO with group filtering of unresolved grades; learning rate see config.
- Weights: bf16. Only the language model is trained; the vision tower (
model.visual.*) and the multi-token-prediction head (mtp.*) are copied unchanged from the initializer (shards named model-untrained-*).
Branches
Table with columns: branch, Megatron iteration, trained tensors, carried over unchanged| branch | Megatron iteration | trained tensors | carried over unchanged |
|---|
checkpoint-10 | iter_0000009 | 427 | 348 |
checkpoint-20 | iter_0000019 | 427 | 348 |
checkpoint-30 | iter_0000029 | 427 | 348 |
checkpoint-40 | iter_0000039 | 427 | 348 |
checkpoint-50 | iter_0000049 | 427 | 348 |
checkpoint-60 | iter_0000059 | 427 | 348 |
checkpoint-70 | iter_0000069 | 427 | 348 |
checkpoint-80 | iter_0000079 | 427 | 348 |
checkpoint-90 | iter_0000089 | 427 | 348 |
checkpoint-100 | iter_0000099 | 427 | 348 |
checkpoint-110 | iter_0000109 | 427 | 348 |
checkpoint-120 | iter_0000119 | 427 | 348 |
checkpoint-130 | iter_0000129 | 427 | 348 |
Eval during training (sair-stage2 205 problems × 4 samples, score/7)
Table with columns: rollout, overall, TRUE problems, FALSE problems| rollout | overall | TRUE problems | FALSE problems |
|---|
| 0 | 0.381 | 0.530 | 0.242 |
| 20 | 0.488 | 0.626 | 0.359 |
| 40 | 0.482 | 0.650 | 0.325 |
| 60 | 0.499 | 0.705 | 0.306 |
| 80 | 0.503 | 0.727 | 0.295 |
| 100 | 0.499 | 0.738 | 0.275 |
| 120 | 0.486 | 0.744 | 0.245 |
Eval points are taken at rollout ids (the weights in use are at most one update behind). Conversion: miles-upstream/tools/convert_torch_dist_to_hf.py functions with the parser namespace rebuilt from the recorded training command; each branch carries conversion_manifest.json with per-shard sha256 and the max |Δ| against the initializer and the previous checkpoint.