Add README

README.md

@@ -0,0 +1,52 @@
+---
+license: apache-2.0
+language: en
+tags:
+  - draft-refine
+  - block-diffusion
+  - nce
+  - qwen3-4b
+---
+
+# Qwen3-4B Stage-3 NCE — beam-lastround variant
+
+Stage-3 Noise-Contrastive-Estimation training resumed from the Stage-2
+end ckpt at step 7335. NCE phase trains the scorer head to rank K=4
+candidate completions per block, with beam-bayes proposal sampling but
+**lastround** score aggregation (only the final K_inner-th iter score
+is used for picking, vs `perround` which averages across iters).
+
+## Files
+
+| File | Size |
+|---|---|
+| `model.pt` | 8.5 GB |
+| `optimizer.pt` | 0.45 GB |
+| `scheduler.pt` | 1.7 KB |
+| `eval_batches.pt` | 13 MB |
+| `rng_rank{0..23}.pt` | 14.7 KB each |
+
+Total: ~9.54 GB / 30 files. Full resume state for re-training.
+
+## Step + lineage
+
+- Resume from: Stage-2 ckpt at step 7335
+- This ckpt: step 10433 (3098 NCE-phase steps; ~2× faster training than
+  perround due to scoring only the final iter)
+- Backbone: Qwen3-4B-Base (frozen during NCE)
+- Config: `configs/large_scale/qwen3_4b_stage3_nce_resume7335_beam_lastround_6n.yaml`
+
+## Eval (Stage-3 BoN with this ckpt)
+
+Audit on gsm8k 200q at α=0.0 (matches Stage-2 baseline):
+| | acc |
+|---|---:|
+| α=0.0 + beam_bayes argmax | **84.00%** |
+
+Identical to perround chain at α=0 (within noise) — both chains' backbones
+remain intact under the freeze. The lastround variant trains 2× faster.
+
+## Related archives
+
+- `haotiansun014/qwen3-4b-stage3-nce-7335-perround-archive`
+- `haotiansun014/qwen3-4b-stage3-nce-7335-temp07-archive`