This project aims to train soft tokens and a backbone VLM to detect reward hacking in target VLMs. But this method may be applied to any multi-modal based retrieval scenarios under flexible computation budget.
Highlights:
-
Data synthesized from carefully designed pipeline of VLM reinforcement learning based finetuning.
-
Solved a few challenging issues brought by the training of soft tokens + LoRA on a larger VLM backbone.
-
Experiments done on multiple VLM series under various parameter sizes including Qwen3-VL and LLaVA-OneVision-1.5.
-
The following figures are from the training of a Qwen3-VL-8B backbone.
-
Setup:
- pip install -r requirements.txt
- install flash attention
-
training:
export HF_TOKEN=<YOUR_HF_TOKEN>; python train.py --model-name Qwen/Qwen3-VL-8B-Instruct --dataset josephzhong/mm-geometry-RewardHacking --problem-col-name question --attn-implementation flash_attention_2 --groups-q 1 2 4 8 16 --groups-c 1 2 4 8 16 --num-epochs 4 --max-length 12288 --batch-size 4 --lr 1e-4 --temperature 0.7 --lora-r 32 --lora-alpha 64 -
evaluation:
python eval.py --model-name <BACKBONE_MODEL_NAME> --save-path <THE_MODEL_SAVE_PATH> --problem-col-name question --dataset josephzhong/mm-geometry-RewardHacking --attn-implementation flash_attention_2 --threshold-precision 0.05 --batch-size 4 --max-length 12288 --temperature 0.7
