prime-grokking.git/src, branch main
Prime Grokking — can minimal architectures grok next-prime? Pre-registered negative-class experiment (weight-tied RNN vs transformer).
http://git.jayrup.me/c/prime-grokking.git/atom?h=main
2026-08-30T21:04:02Z
fix(eval): structured-mode greedy decode capped at in-range layout bound (369) — probe-window traces reach 633 tokens and were silently truncated; cap now keys to positional table (1024) so OOD traces decode fully. Regression-verified vs stub model.
2026-08-30T21:04:02Z
Void Agent
void@jayrup.hermes
2026-08-30T21:04:02Z
urn:sha1:3135fea726157f8ea596758ce246dfb74779c865
fix(transformer): increase positional embedding and causal mask buffer to 1024 to support worst-case probe traces
2026-08-29T22:35:10Z
CaptainJack2491
jayrupnakawala@gmail.com
2026-08-29T22:35:10Z
urn:sha1:fc56dd007142264ef029eb8c0f9db11a6cf3d298
optimize E8: precompute causal mask buffer, add eval early exit, and enable compile_model in e8.csv
2026-08-29T22:32:30Z
CaptainJack2491
jayrupnakawala@gmail.com
2026-08-29T22:32:30Z
urn:sha1:9a5912f7764dbc942b56f8114f2c878c1a6d44ff
implement Addendum 8: 6-arm token-space recurrence suite (arms A, B, C, D1, D2, D3), tests, and jobs/e8.csv
2026-08-29T22:19:29Z
CaptainJack2491
jayrupnakawala@gmail.com
2026-08-29T22:19:29Z
urn:sha1:7b0fddb02b089b82b8d12b2dcac17bf9527817da
eval: strip torch.compile _orig_mod./module. prefixes when loading checkpoints (salvages compiled-GPU runs)
2026-08-24T23:33:03Z
Void Agent
void@jayrup.hermes
2026-08-24T23:33:03Z
urn:sha1:4b0c3cdc3db5b29c9d770082bf6caeb677221bfa
fix(train): enforce epoch permutation without replacement, add max cache OOM guard, and handle empty split tensors
2026-08-20T18:25:57Z
CaptainJack2491
jayrupnakawala@gmail.com
2026-08-20T18:25:57Z
urn:sha1:615c87402138da053abe2cd0ab897eb6a5af9d67
perf(train): VRAM-resident tensor caching, zero-copy GPU batch sampling, and adaptive eval schedule
2026-08-20T16:37:46Z
CaptainJack2491
jayrupnakawala@gmail.com
2026-08-20T16:37:46Z
urn:sha1:be8d62351497ed2e5ee136ae533ee9c5ec4491f2
fix(eval): correct sieve-rank indexing + test expectations
2026-08-20T16:02:15Z
Void Agent
void@jayrup.hermes
2026-08-20T16:02:15Z
urn:sha1:cf83937689bb30e2b5fa6e3efaa2f115016a030b
- _compute_sieve_rank: find LARGEST k (not smallest) whose signature
contains all error preds — fixes P5(4) vs P5(1) bug
- _sieve_rank_signature: iterate all small_primes (not just p²≤n) to
catch composites like 209=11*19 where 11²>209
- Test expectations updated for [101,200] range (209 outside it)
- is_prime task: P1→P3 (no sieve-rank for classification tasks)
48/48 green.
fix(eval): P5/P6 sieve-rank ladder replaces hardcoded FLAGGED_SIEVE_PREDS
2026-08-20T15:55:49Z
Void Agent
void@jayrup.hermes
2026-08-20T15:55:49Z
urn:sha1:f8995751de925571d4eec669ab7ab8b1c3d76cb9
- probe_report: dynamic lo/hi (no more hardcoded [101,200])
- main(): probe range derived from range_end ([range_end+1, range_end+1000])
- P6 = exact (no probe misses), P5(k) = errors match rank-k sieve signature
- _sieve_rank_signature: computes composites with all factors > p_k
- _compute_sieve_rank: identifies rank from model's error predictions
- build_report.py: probe_fig uses sieve rank for coloring
- Tests updated for P5/P6 codes (48/48 green)
perf(cuda): add AMP fp16 autocast with GradScaler and torch.compile reduce-overhead mode
2026-08-17T15:35:28Z
CaptainJack2491
jayrupnakawala@gmail.com
2026-08-17T15:35:28Z
urn:sha1:7166dc47c9bd764e2cd1d7c3d9e046cdc439b58c