weixu-zhang/viscoder2-3b-grpo-phase3d-7lang-step1500

VisCoder2-3B GRPO Phase 3d 7-language step 1500

This is the HF-merged model for the latest clean checkpoint from the H200 Phase 3d 7-language VisRL run.

Provenance

• —Base model: TIGER-Lab/VisCoder2-3B
• —Training method: GRPO with rubric rewards
• —Training data: multilang_7lang_with_rubrics_in_extra.parquet (6999 rows)
• —Trainable languages: Python, HTML, SVG, Vega-Lite, Mermaid, LilyPond, LaTeX
• —Judge used during the H200 run: Qwen/Qwen2-VL-7B-Instruct
• —Clean checkpoint: global_step_1500
• —Raw run progress: training continued to logged step 1577, but 1500 is the last saved safe checkpoint

What this repo is for

Use this repo for:

• —vLLM / Transformers inference
• —bench evaluation on another server
• —qualitative inspection of the current best H200 Phase 3d policy

If you want to resume training, use the raw FSDP checkpoint dataset instead: weixu-zhang/visrl-resume-checkpoints.

Notes

• —This is an accepted partial checkpoint, not an end-of-training validated final model.
• —The corresponding experiment log lives in docs/H200_EXPERIMENT_LOG_2026-05-02.md in the source repo.