Experiments
-
DFlash Qwen3-0.6B — batch 24, 10 epochs success
The 6-epoch run was underfitting, so this doubles per-GPU batch to 24 (global 192), extends to 10 epochs, and sqrt-scales the learning rate to 8.5e-4.
- Experiment description
- GPU server
- Experiment preparation
- Sanity checks
- Full training
- Full evaluation
-
DFlash Qwen3-0.6B — 1024 len, batch 12, 6 epochs success
First full 1024-length DFlash run on the fast-iteration setup. Reached val/greedy_accepted_length 4.770; validation was still rising at the end.
- Experiment description
- GPU server
- Experiment preparation
- Sanity checks
- Full training
- Full evaluation