VisCoder2-3B GRPO Phase3e 8lang 3k/lang (step 5721)
Merged Hugging Face export of the final checkpoint from the phase3e full-parameter
3B run trained on the 8-language 3k/lang clean bundle.
- Base model:
TIGER-Lab/VisCoder2-3B
- Training recipe: GRPO full-parameter finetuning
- Data:
- Python: legacy balanced
python_3000
- Other 7 languages: HF-cleaned 3k/lang bundle
- LaTeX: HF-clean 2877 + 123 legacy top-up
- Languages: python, html, svg, vegalite, mermaid, lilypond, latex, asymptote
- Final checkpoint:
global_step_5721
Local source run:
- Output dir:
viscoder2_3b_grpo_phase3e_8lang3k_hfclean_legacypy_20260515_0942
- Merged actor:
global_step_5721/actor/hf_merged
This artifact is intended for evaluation / inference. Resume training should use
the raw FSDP checkpoint rather than this merged export.
