ViRL39K 7B RL arms (Qwen2.5-VL-7B-Instruct, unfrozen vision tower)
Research checkpoints from a controlled multi-arm RL study. Each folder is a merged HF checkpoint at global_step_162 (2 epochs, seed 1, unfrozen vision tower).
| folder | arm | objective |
|---|---|---|
A1-7B-uvit/ |
A1 | GRPO, binary-accuracy reward |
A3-7B-uvit/ |
A3 | VPPO (visual-dependency credit assignment) |
A4-7B-uvit/ |
A4 | PRPO + RVD |
A0 (base) is not here: it is unmodified Qwen/Qwen2.5-VL-7B-Instruct. Pull it from there.
Shared recipe: ViRL39K (Geo3K-deduped), 162 steps, global batch 128, rollout 384 x n8, lr 1e-6, DAPO clip 0.2/0.28, token-level loss, max response 2048.
Fetch one arm:
from huggingface_hub import snapshot_download
snapshot_download("Ahsanz/virl39k-qwen25vl-7b-arms", allow_patterns="A1-7B-uvit/*", local_dir="arms")