Add run2/run3 one-shot missing-anchor caches (Qwen3-4B 20k N=3)

Publish Qwen3-4B subset20k anchor data and generation guide

Add if_ref_anchor_teacher4b_reasoning_train_seed1_val512_t1_p095_k20_pp0_r8192_plus32768fallback_run3_scored_by_qwen3_4b.json (interim: stage B 11264/16585 rows)

Add if_ref_anchor_teacher4b_reasoning_train_seed1_val512_t1_p095_k20_pp0_r8192_plus32768fallback_run2_scored_by_qwen3_4b.json (interim: stage B 11264/16585 rows)

FINAL 4096-row experiment subset: 3620 r8192 rows + 476 targeted r32768 fills; loader-verified complete; provenance in repo data_subsets/

Add if_ref_anchor_teacher4b_reasoning_train_seed1_val512_t1_p095_k20_pp0_r8192_plus476r32768_scored_by_qwen3_4b.json (476 v2 anchor rows refilled at 32768-token budget)

Add regenerated v2 Qwen3-4B reasoning anchor cache (t1.0/p0.95/k20/pp0.0, 8192-token budget, no empty-response retries). 93882 rows: 82513 complete, 76691 hygiene-clean. Differs from the base v2 file in that the 21994 rows failing the hygiene validator were deleted and redrawn once (repetition loops -60%/-68%, near-empty answers largely persistent).

Add v2 Qwen3-4B reasoning anchor cache (t1.0/p0.95/k20/pp0.0, 8192-token budget). 93882 rows, 81044 complete; single build pass, no empty-response retries.

Add validated Tulu-3 8B DPO anchors (1, 2048 tokens)

Add Llama-3.1-8B-Instruct non-reasoning anchor cache (y0/y1 + ref PPL, seed=1, val=512, scored by Llama-3.1-8B-Instruct)

Replace Qwen3-4B nonreason anchor cache with val512 cache

Upload Qwen3 8B anchor cache

Upload OLMo3 7B anchor cache

Add files using upload-large-folder tool

Upload Qwen3-4B nonreason anchor cache

Upload if_ref_anchor_teacher4b_reasoning_train_seed1_scored_by_qwen3_4b.json

Upload teacher17b nonreason anchor cache

Upload if_ref_anchor_teacher17b_reasoning_train_seed1_scored_by_qwen3_17b.json

initial commit