kent/consciousness
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
consciousness/training
History
ProofOfConcept ac9a9034fb apollo: rewrite optimizer from paper's math + add research analysis 
Corrections from reading the full paper (arXiv:2412.05270):
- Add gradient scale factor α = √(n/r) — compensates for systematic
  ratio between compact and original space scaling factors
- Add norm-growth limiter (γ=1.01) — prevents loss spikes in early training
- Refresh projection matrix every 200 steps, not every step
- Channel-wise scaling for rank>1, tensor-wise for rank=1
- Scaling applies as G·diag(s), preserving gradient direction per channel

Research writeup in training/research/apollo-paper-analysis.md covers:
- Full mathematical derivation (equations 1-9)
- Theorems 4.1 and 4.2 (JL-based approximation bounds)
- Why Apollo can beat AdamW (directional sharpness, Hessian spectra)
- Fine-tuning results (matches AdamW at 0 memory cost)
- Ablation studies (rank, scaling granularity, projection method)
- Implications for our behavioral fine-tuning use case
2026-03-31 00:54:17 -04:00
..
checkpoint checkpoint: sync live weights back into model safetensors in-place 2026-03-30 22:55:23 -04:00
research apollo: rewrite optimizer from paper's math + add research analysis 2026-03-31 00:54:17 -04:00
apollo_mini.py apollo: rewrite optimizer from paper's math + add research analysis 2026-03-31 00:54:17 -04:00
apollo_worker.py apollo: make rank configurable (default 1 = Mini, higher ranks for experimentation) 2026-03-30 22:06:31 -04:00
DESIGN.md DESIGN.md: complete rewrite reflecting validated architecture 2026-03-31 00:42:53 -04:00
export_weights.py apollo-mini training system: initial implementation 2026-03-30 22:02:37 -04:00
start_vllm_with_apollo.sh vllm launcher with apollo hook 2026-03-30 22:24:02 -04:00
train.py apollo-mini training system: initial implementation 2026-03-30 22:02:37 -04:00
training_example.py apollo-mini training system: initial implementation 2026-03-30 22:02:37 -04:00
vllm_export_hook.py apollo-checkpoint: efficient diff-based GPU weight checkpointing 2026-03-30 22:53:17 -04:00
weight_mapping.py weight_mapping: strip language_model prefix to match HF text model names 2026-03-30 23:11:03 -04:00