VisRL 7B LoRA low-LR step300
This repo stores the merged HF checkpoint for the current best single-H200
7B LoRA run in the VisRL project.
What it is
- base model:
TIGER-Lab/VisCoder2-7B
- training recipe: GRPO + LoRA on the Phase 3b fixed-rubric mixed-language data
- judge during training:
Qwen/Qwen2-VL-7B-Instruct
- hardware: single NVIDIA H200
- branch: low-LR continuation from the original LoRA
global_step_200
- learning rate on this branch:
2e-6
- selected checkpoint: local
global_step_300
Why this checkpoint was selected
On the generation-only fuller VisPlotBench comparison (python196 + svg65),
this checkpoint was the best 7B result seen so far in the repo:
| model | python clean | python has_plot | python exec_err | python incorrect | svg clean | svg has_plot | svg exec_err | svg incorrect |
|---|
| base VisCoder2-7B | 126 | 175 | 70 | 21 | 64 | 64 | 1 | 1 |
this checkpoint (step300) | 128 | 176 | 68 | 20 | 64 | 64 | 1 | 1 |
A follow-up tail continuation from this sweet spot with LEARNING_RATE=1e-6
produced clean local step50 and step100 checkpoints, but neither beat this
step300 checkpoint overall:
| checkpoint | python clean | svg clean |
|---|
1e-6 step50 | 128 | 63 |
1e-6 step100 | 126 | 64 |
Use this repo when
- you want the current best 7B checkpoint for eval / inference / bench runs;
- you want a merged HF-format artifact rather than raw FSDP shards;
- you are migrating to a new server and need a pullable model repo.
If you want to resume training
Use the raw checkpoint mirror in:
- dataset repo:
weixu-zhang/visrl-resume-checkpoints
- path:
global_step_300/
