wassname/ml-debug
1
0
Fork 0
mirror of https://github.com/wassname/ml-debug.git synced 2026-06-27 16:00:43 +08:00
Code Issues Packages Projects Releases Wiki Activity
62 Commits 1 Branch 0 Tags
main
Commit Graph

9 Commits

Author SHA1 Message Date
wassname b8c3ffcf11 gpt5.5/fable 2026-06-12 09:30:25 +08:00
wassname 8b9a1d62ed docs: resolve ml-debug TODO references 2026-06-12 06:52:38 +08:00
wassname (Michael J Clark) 6e9a3ca633 Revise introduction in diagnostics.md 
Updated the introduction to provide context for diagnostic code snippets.
 2026-06-11 21:18:07 +08:00
wassname 8ee980d62f diagnostics: add NaN-poisoning leakage tracer + Karpathy backprop-to-input check; README citation 
NaN poisoning: inject NaN where info must not come from (future/test/labels), run the real pipeline, assert past outputs stay finite. Documents false negatives (pandas skipna, nanmean) and false positives (softmax rows, batch stats). Backprop-to-input is its gradient dual for inside the model; quote already frozen in docs/evidence/karpathy_recipe_training_nn_2019.md.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
 2026-06-11 10:18:51 +08:00
wassname c9c53f8e7f feat(ml_debug): expand nanochat evidence, add lec4 diagnostics file 
nanochat_deepwiki_llm_pretraining_2026.md rewritten with content from
dev/LOG.md and deepwiki sections 3/12/13:
- 14 labelled findings with direct quotes and empirical numbers
- Dataset >> architecture (27% gain, 5 failed attempts before ClimbMix)
- Scale-dependent HP sensitivity (d12 HPs hurt d20)
- Multi-axis validation (steps/wall-clock/FLOPs)
- Negative results: MoE/SwiGLU/MTP all failed at this scale
- MFU monitoring, batch size Bopt∝D^0.383, WD∝1/width² tables
- FP8 reality: 1.38x micro → 1.17x full → 5% capability-matched
- Python GC 500ms overhead, torch.compile recompile gotcha

karpathy_nn_zero_to_hero_lec4_diagnostics.md: new evidence file
- Activation saturation check (tanh >0.97)
- Gradient distribution check per-layer
- Grad:data ratio (target ~1e-3)
- Update-to-data ratio tracker with full plotting code
- Incremental improvement log from notebook
 2026-03-10 05:38:33 +08:00
wassname ced4edc200 feat(ml_debug): add Karpathy recipe + nanochat evidence, update-ratio diagnostic 
Add 3 new evidence files from modern open-source sources:
- karpathy_recipe_training_nn_2019.md: Karpathy's training recipe blog post
- nanochat_deepwiki_llm_pretraining_2026.md: 320+ HP sweeps for GPT-2-scale pretraining
- sanh_simple_considerations_hf_2021.md: HuggingFace NLP debugging notes

Add update-to-data ratio diagnostic to refs/diagnostics.md (target ~1e-3).
Add LLM pretraining gap note to SKILL.md intro linking the new sources.
Add tanh saturation % to logging checklist.
 2026-03-10 05:32:37 +08:00
wassname bbe3fe0985 feat(ml_debug): add JAX grep patterns and diagnostic equivalents 
refs/static_analysis.md: JAX-specific grep patterns (in-place mutation,
print side effects, key reuse, numpy escape, cast behavior).
refs/diagnostics.md: JAX equivalents table (NaN detection, gradcheck,
disable_jit, debug.print, debug.breakpoint, checkify).
 2026-03-06 14:10:39 +08:00
wassname 7ac7aacac7 fix(ml_debug): address review feedback 
- Fix stale Part 2 cross-references to link to rl/SKILL.md
- Add McCandlish + Slavv back to parent Sources (cited in Part 7)
- Add back-links from refs/ files to parent SKILL.md
 2026-03-06 13:59:48 +08:00
wassname 70c28f06ac refactor(ml_debug): extract grep patterns and diagnostics to refs/ 
Moved 6.1 (static analysis grep patterns) and 6.2 (diagnostic code
snippets) to refs/static_analysis.md and refs/diagnostics.md.
Triage tree (6.3) stays in main with references to the ref files.
ml_debug/SKILL.md reduced from 7229w to 5093w (~30% from original).
 2026-03-06 13:54:37 +08:00