Nils - Gitea: Git with a cup of tea

nils pushed to main at nils/jspace

2026-07-14 23:09:00 +00:00

3477fb5da0 method: truncated-BPTT bias bound via rho(A) — contraction certifies tail-only gradients

nils pushed to main at nils/jspace

2026-07-14 23:07:03 +00:00

d00f120f27 PerDepthAdapter (Bae-style per-iteration merges) + convergence-halting probe (free ACT); pre-registration item 13

nils pushed to main at nils/jspace

2026-07-14 22:53:37 +00:00

f254b20b62 ParcaeAdapter: rho(A)<1 by construction (ZOH negative-diag), rho logging in rec arms, pre-registration item 12

nils pushed to main at nils/jspace

2026-07-14 22:49:46 +00:00

d2da8044c3 RecurrentAdapter arm: Huginn-regime retrofit (learned A/B, noise h0, randomized depth) + pre-registration item 11

nils pushed to main at nils/jspace

2026-07-14 18:09:00 +00:00

0fd93cb328 n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized

nils pushed to main at nils/jspace

2026-07-14 17:44:10 +00:00

1d4cffd6ad integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

caef95f10f budget-CoT per-item: paired ties with latent arms confirmed (p>=0.69)

2ce646d175 best-of-3 oracle/deployable split: realistic selector loses the overall edge (55.0/27.3 vs oracle 57.8/34.5); paper economics rewritten

e5fc453005 pin gemma-4 model revisions (upstream 12B template regression: thought channel -> empty generations)

Compare 4 commits »

nils pushed to main at nils/jspace

2026-07-14 14:56:19 +00:00

378bb36b8c node_setup: retry+timeout around model downloads (xet mid-transfer stalls)

nils pushed to main at nils/jspace

2026-07-14 14:48:15 +00:00

7821f304f8 LiveCodeBench transfer harness: raw-jsonl loader (script datasets dead), stdin judge, newest-window selection, plan labeling

nils pushed to main at nils/jspace

2026-07-14 14:30:37 +00:00

dcd0b0b2dc HumanEval transfer: trained-vs-untrained-merge paired test (p=0.80, n.s.) -- transfer gain is the merge itself, not trained content

nils pushed to main at nils/jspace

2026-07-14 14:05:41 +00:00

d1bafc2d4d revision per review: seed means in headlines, net accounting vs untrained merge, bucket definition up front, GSM12B n.s. (p=0.86), law->regularity, norm spec, compute appendix, public repo URL

nils pushed to main at nils/jspace

2026-07-14 14:00:20 +00:00

01fd272028 panel: chat-template MC scoring (raw scoring is chance on -it model), parquet dataset mirrors, incremental dump

nils pushed to main at nils/jspace

2026-07-14 13:48:38 +00:00

87308cfb0b score outcomes against pre-registered protocol items 1-10

nils pushed to main at nils/jspace

2026-07-14 13:47:18 +00:00

55665ce251 CPU endgame: stats pass (Wilson/McNemar/pooled hard), final figures, PAPER.md rewrite around amortizable-content thesis

nils pushed to main at nils/jspace

2026-07-14 11:59:00 +00:00

6a954bd3e4 MC capability panel: ARC/WinoGrande/HellaSwag/MMLU with prompt-loop

nils pushed to main at nils/jspace

2026-07-14 11:02:49 +00:00

b461970c9a warm-start flag (stacking arm), adaptive flags for Blocksworld

nils pushed to main at nils/jspace

2026-07-14 11:00:39 +00:00

aa3e11fb20 gpuq worker: tee job output to pane for live viewing

nils pushed to main at nils/jspace

2026-07-14 10:57:10 +00:00

f2cbce5a03 related_work: log 12b_adaptive outcome as confirmation of the anchoring account

nils pushed to main at nils/jspace

2026-07-14 10:50:46 +00:00

e152a91c05 parse_known_args: module-level parser must tolerate importers' argv

nils pushed to main at nils/jspace

2026-07-14 10:47:25 +00:00

cd8e94dda5 gpuq: idle-time class dispatcher (single-writer, race-free) + docs

nils pushed to main at nils/jspace

2026-07-14 10:41:36 +00:00

9329f22a63 gpuq: class-based scheduling at submit time (pending count + live util)