label-free-bias-steering-checkpoints

Label-Free Bias-Only Steering — checkpoints (deprecated layout)

This repo is kept for archival purposes. The checkpoints below have been split into one repo per benchmark/base model, each with its own model card and a standalone load_bias.py loading script:

• label-free-bias-steering-ai2d-qwen2.5-vl-7b
• label-free-bias-steering-logicvista-qwen2.5-vl-7b
• label-free-bias-steering-mathvista-qwen2.5-vl-7b
• label-free-bias-steering-mmau-qwen2.5-omni-7b
• label-free-bias-steering-math500-qwen2.5-7b
• label-free-bias-steering-math500-qwen2.5-math-7b

New usage/results should reference those repos, not this one.

Bias-only steering vectors trained with label-free test-time reinforcement learning (majority-vote pseudo-label reward, GRPO), one per benchmark. Each checkpoint is ~100K bf16 parameters — a per-layer additive bias vector for the corresponding frozen backbone. See the paper for method details.

Files

Each file is a single representative training seed; the paper reports mean ± std over 3 seeds per benchmark, so these numbers vary slightly from the paper's headline figures.

Usage

Each checkpoint is a plain torch.save'd state_dict-style mapping from parameter name (e.g. model.layers.N.mlp.down_proj.bias) to tensor. Load with torch.load(path) and copy into the matching frozen backbone — or use the per-benchmark load_bias.py scripts in the split repos linked above, which are self-contained.