fengyikang/internvl2-5-2b-rm-v1

InternVL2.5-2B Reward Model (v1, Linear head)

Discriminative scalar reward model trained on VLFeedback (Bradley-Terry), evaluated on VL-RewardBench test.

Metrics (pointwise + K=5)

Overall	Macro	agree	reject	tie
53.34%	49.64%	567	496	184

Files

• model.safetensors: fine-tuned InternVL backbone
• score_head.pt: Linear score head
• reward_config.json: head config (if present)

Usage

Requires tokenizer/model code from OpenGVLab/InternVL2_5-2B.

python eval_reward_vlrewardbench.py \
  --checkpoint-dir ./ \
  --base-model-path OpenGVLab/InternVL2_5-2B \
  --scoring pointwise --k 5