Ckpt Archive 20260429 Public

# Draft-Refine training-ckpt archive (2026-04-29) Latest step-numbered checkpoint of three concurrent training chains. ## Subdirectory mapping | Tag | Chain (descriptor) | Output dir on cluster | |---|---|---| | `n1_*` | Qwen3-4B Stage-3 NCE pool-mlp-beam (frozen-baseline scorer) | `outputs/qwen3-4b-stage3-nce-pool-mlp-beam-resume4000-4n` | | `b8_*` | Qwen3-8B-Base CPT on 10B-tok corpus, σ ramp 0.70→0.90, K_max=3 unified Stage 1+2 | `outputs/qwen3-8b-base-cpt-10b-sigma070-090-warmup-8n-kmax3` | | `s4_*` | SDAR-4B-Chat SFT on 10B-tok, σ ramp 0.65→0.85, K_max=3 unified Stage 1+2 | `outputs/sdar-4b-warmup-sft-sigma065-085-8n-kmax3` | | `b4_*` | Qwen3-4B-Base CPT on 10B-tok corpus, σ ramp 0.70→0.90, K_max=3 unified Stage 1+2 (Bristen A100) | `outputs/qwen3-4b-base-cpt-10b-sigma070-090-warmup-bristen-8n` | Each tag's directory contains `model.pt` (full DDP state_dict, including backbone + scorer if present), `optimizer.pt`, `training_state.json`, `eval_batches.pt` (persisted eval batches for resume), and per-rank RNG. ## Upload script `scripts/hf_upload_public_2026-04-29.py` (in the draft-refine repo).