LLaDA-8B full-sequence SFT with code-target continuation
Matched full-sequence-SFT code control for the dLLM Registers paper. This checkpoint has no register or discrete-text carry channel.
Training
- Initialization:
albertge/llada-8b-full-sft-mix60k-4pass - Continuation data: the 30K OpenCodeInstruct subset of
albertge/mix60k-math-code-sft - Targets: raw Python delimited by
<code>...</code> - Four independently noised diffusion-loss optimizer passes per example
- Natural/dynamic sequence lengths, capped at 1024 completion tokens
- One epoch, batch size 1 per rank, learning rate 1e-5, weight decay 0.1
- Exact runner:
SFT/sft_train_vanilla.pyandmodal_jobs/vanilla_sft.py
Recorded bounded evaluation
Fresh 16 x 64 clears generated text at every boundary and carries no state. Chunks are concatenated and scored as one program.
| Protocol | HumanEval | MBPP |
|---|---|---|
| Fresh 16 x 64, no carried state | 11.6 (19/164) | 23.7 (61/257) |
Repository: https://github.com/lbertge/d1-registers