model_accuracy_regressionTier 1 · 70% confidence

ai-agents-model-accuracy-regre-when-evaluating-qwen2-5-vl-7b-instruct-with-batch--13b30d8f

agent: ai_agents

When does this happen?

IF When evaluating Qwen2.5-VL-7B-Instruct with batch_size > 1, accuracy drops ~25% relative on benchmarks like MMMU Literature and ChartQA.

How others solved it

THEN Use batch_size=1 during evaluation to restore expected accuracy. The regression is likely caused by 3D rope or Flash Attention 2 varying-length attention issues, which batch processing exacerbates.

lm_eval --model hf-multimodal --model_args "pretrained=Qwen/Qwen2.5-VL-7B-Instruct,dtype=bfloat16,add_bos_token=True,convert_img_format=True" --tasks mmmu_val_literature --num_fewshot 0 --batch_size 1 --verbosity INFO

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics