ttc-trunk-8b-r2048-33k — the TTC headliner (8B, state rank 2048)

Main representative of the TTC cache writer (Neehal, 2026-09-24): Qwen/Qwen3-8B backbone (frozen), state rank 2048 rows, 33K-token streams (block 2048, FIFO + 3 sinks, fold rowmax), T data = ProLong .8 / generated doc-QA v2 .2; 6 nodes x 48 streams per optimizer step, lr 1e-3 constant (WSD trunk, no in-run decay), 4,000 steps planned (= the 194k-stream budget; training continues). Memory accounting at this rank: state_tok_eq_measured = 2096 KV-token equivalents (KeyDiff comparison total 4147 at the 2051 budget).

• branches/a1000_mixB/writer_s001250.pt: the a1000 anneal — 250-step arm-B decay from the step-1000 retained save (lr 1e-3 -> 0, 8 streams/step; mixture text .1 / SQuAD .6 / TTL .2 / NIAH-multiquery .05 / VT .05), stored step 1250, optimizer stripped (writer, shared, args, step). With args.json, train_log.jsonl, branch.log, readouts.log and results/: the 20 stitched-grid cells base_<bed>_na_ttc8b_r2048_a1000_c<3|7|11|15>_s0.json (squad / tqa / niah_single / niah_multikey / vt, arms full_kv / stream_ttc / stream_ttc_wrong, n=50, block 2048, seed 0), the whole-document set (*_paper_*), the 16K / 33K / longmino / prolong ladders and LongPPL.
• trunk/step001000/ckpt_step001000.pt: the raw (undecayed) trunk at step 1000 the anneal started from, with its readouts and the same 20 grid cells under grid/ (tag ttc8b_r2048_s1000) for the like-for-like.

Anneal vs raw snapshot (stream_ttc, evicted-state site, 20 cells): mean +.160; NIAH single +.25 to +.36 and multikey +.20 to +.29 at c7-c15, VT +.17 to +.47, SQuAD +.03 at c3/c7, TQA +.05 at c3/c11. Language ladders move by ~.00 KL / -.02 key_closed.

Evaluate with the repo's python -m eval.eval_ruler_stream --model Qwen/Qwen3-8B --ckpt <pt> --task <task> --tqa-bed data/beds/tqa_stream_bed_qwen3-8b_v2.jsonl --cycles <c> --block-tokens 2048 --n 50 --seed 0 --num-sinks 3 --state-in-prefill on --stream-mode documents --filler auto --vocab-split heldout --fold rowmax --pos-mode absolute --arms full_kv stream_ttc stream_ttc_wrong --kl-vs-full (env SUITE_QA_BACKBONE=qwen3-1.7b SUITE_MK3_KEYS=words NEEDLE_VOCAB_SPLIT_VERSION=1). Later anneals (a2000, final) land under branches/. Uploaded by Neehal's session on 2026-09-25; the r8192 lane is in ntumm120/ttc-trunk-8b-r8192-33k.

Update 2026-09-25 (the trunk continues to 4,000 steps = the data end; Neehal 2026-09-24)

• branches/a2000_mixB/writer_s002250.pt: the a2000 anneal — 250-step arm-B decay from the step-2000 retained save (same mixture, lr 1e-3 -> 0, 8 streams/step), stored step 2250, optimizer stripped; with args.json, train_log.jsonl, branch.log, readouts.log, results/ (whole-document set, ladders; the stitched grid cells follow when run).
• trunk/step002000/ckpt_step002000.pt: the raw trunk at step 2000 with its readouts.