fengyikang/internvl2-5-2b-rm-v2-oversample

InternVL2.5-2B Reward Model (v2, MLP, hard oversampling)

Same v2 recipe as the no-oversampling variant, with hard pair oversampling (weight=2). Training indices: 70,761.

Metrics (pointwise + K=5)

Overall	Macro	agree	reject	tie
66.12%	62.17%	765	392	90

Usage

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