ViRL39K 3B RL arms — FROZEN vision tower
Research checkpoints from a controlled multi-arm RL study on ViRL39K, at global_step_162 (seed 1). Every arm here was trained with the vision tower frozen; the trainable-tower siblings are published separately (see below).
| folder | objective | local name in the study |
|---|---|---|
A1-3B-frozen |
GRPO, binary-accuracy reward | A1 |
A3-3B-frozen |
VPPO (visual-dependency credit assignment) | A3 |
A4-3B-frozen |
PRPO + RVD | A4-frozen |
Frozen is verified, not asserted. Each folder's 390 visual.* tensors hash to df29608efbaa81d1, bitwise
identical to the stock Qwen/Qwen2.5-VL-3B-Instruct
vision tower. The trainable siblings hash differently.
Naming warning
In the study's own files the 3B naming is inverted relative to the other backbones: the bare A1,
A3 and A4-frozen are the FROZEN arms, while A1-uvit, A3-uvit and A4 are the TRAINABLE
ones. At 7B it is the other way round (A1-7B is trainable, A1-7B-frozen is frozen). The folders
here name the tower explicitly so that trap is not exported.
Trainable-tower siblings
- GRPO:
Ahsanz/e2-virl39k-3b-a1-uvit - VPPO:
Ahsanz/e2-virl39k-3b-a3-uvit - PRPO:
Ahsanz/e2-virl39k-3b-a4
A0 (base) is not here: it is unmodified Qwen/Qwen2.5-VL-3B-Instruct.