 sleepy
|
1dd5fe46a8
|
Results: plateau warmup, lambda=0.5, best PPL 36.0 at lambda=0.05
|
2026-04-24 05:17:59 +02:00 |
|
|
sleepy
|
322316fc5f
|
Add plateau warmup schedule for gradual quantization
|
2026-04-24 04:45:13 +02:00 |
|
|
sleepy
|
c10212735a
|
Fix: use abs_mean scale for balanced ternary distribution
|
2026-04-24 04:15:01 +02:00 |
|
|
sleepy
|
853019baf2
|
Simplify: remove learnable scale, use abs_mean+round, slow warmup
|
2026-04-24 04:10:21 +02:00 |
|
|
sleepy
|
27e9faf4f5
|
Fix: abs_mean+rounding quant, slow warmup (2000 steps), threshold param, lower LR
|
2026-04-24 03:38:37 +02:00 |
|
|
sleepy
|
868910b40f
|
Initial ternary quantization framework: BitLinear, QAT training loop, eval harness
|
2026-04-24 03:07:55 +02:00 |
|
|
sleepy
|
7378d4ef8f
|
Add ternary QAT training pipeline: prepare.py (data/eval), train.py (quantization/training), program.md (agent instructions), autoresearch.sh (loop)
|
2026-04-24 01:36:44 +02:00 |
|
|
sleepy
|
f4601547d2
|
Initial commit: PLAN.md
|
2026-04-24 00:32:55 +02:00 |
|