YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
TTC 8B trunk, state rank 4096 (33K streams) — a2000_mixB anneal
Qwen/Qwen3-8B backbone (frozen), state rank 4096 rows, 33K-token streams (block 2048, FIFO + 3 sinks, fold rowmax), the same T data and trunk recipe as the headliner (ntumm120/ttc-trunk-8b-r2048-33k) except the state size; cache substrate, recurrent write, writer d_c 512, lr 1e-3 constant (WSD trunk), resumable saves every 500 steps. Ladder siblings: r288, r512, r1024, r2048 (headliner), r4096, r8192 (ntumm120/ttc-trunk-8b-r-33k).
- branches/a2000_mixB/writer_s002250.pt: the a2000_mixB anneal, a 250-step arm-B decay (mixB) from the step-2000 retained save (lr 1e-3 -> 0), optimizer stripped (keys writer, shared, args, step); with args.json, train_log.jsonl, branch.log, readouts.log and results/ (the whole-document set where it was run: tqa_contig c3/c7, tqa c3, niah single/multikey contig c7, tqa_doc shards, LongPPL tqa_contig).
Evaluate exactly as described in the r2048 repo's README (same eval command and environment). Uploaded by Neehal's session on 2026-09-25.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support