jspace-nanogpt.git/src/loss_reweight.py, branch main
J-space on nanoGPT: Jacobian Lens replication & critique. J-lens norms anti-correlate with token frequency (r ~ -0.61 to -0.69); frequency-matched synthetic pair + causal loss-reweighting tests.
http://git.jayrup.me/c/jspace-nanogpt.git/atom?h=main
2026-08-02T13:24:51Z
Address Luna repo review: dynamic repo root (no hardcoded cwd), J-lens correctness test (last-layer identity), auditable fact-check, mechanism narrative fix, softened conditional-predictability claim, README data-prep + artifacts note
2026-08-02T13:24:51Z
Void Agent
void@jayrup.hermes
2026-08-02T13:24:51Z
urn:sha1:31e1656f1fd4409c14108e63eb3e430f657746ac
Document n_pairs denominator (exact count) and ctrl_random gradient-mass caveat (comment-only)
2026-08-02T13:20:05Z
Void Agent
void@jayrup.hermes
2026-08-02T13:20:05Z
urn:sha1:2ba0e14c3559e5786c324a89f26f159363a230b5
loss_reweight: pass Y to forward for full logits (Karpathy nanoGPT returns last-position logits when targets=None); regression test
2026-08-02T12:32:16Z
Void Agent
void@jayrup.hermes
2026-08-02T12:32:16Z
urn:sha1:53de9ecd13fbf7e098fea88f73d1a9403a40d83b
loss_reweight: CPU generators for randint/randperm (CUDA generators unsupported on torch 2.4); cuda-path verified
2026-08-02T12:29:24Z
Void Agent
void@jayrup.hermes
2026-08-02T12:29:24Z
urn:sha1:799df091c7a163607cca1d5dcafb664f915ca847
Add causal experiments: synthetic frequency-matched pair (Gemini design) + loss-reweighting (Codex design); jlens_v3 output_dir + last-layer device fix
2026-08-02T11:41:38Z
Void Agent
void@jayrup.hermes
2026-08-02T11:41:38Z
urn:sha1:8c2dc42642b1cdbfb5dfacd9c188821b03c84c4b