dataset_path: text
dataset_kwargs:
  data_files:
    test: https://raw.githubusercontent.com/zai-org/MotionBench/main/data/video_info.meta.jsonl
task: motionbench_full
test_split: test
output_type: generate_until
process_docs: !function utils.motionbench_process_docs
doc_to_visual: !function utils.motionbench_doc_to_visual
doc_to_text: !function utils.motionbench_doc_to_text
doc_to_messages: !function utils.motionbench_doc_to_messages
doc_to_target: "answer"
generation_kwargs:
  max_new_tokens: 64
  temperature: 0
  top_p: 1.0
  num_beams: 1
  do_sample: false
process_results: !function utils.motionbench_process_results
metric_list:
  - metric: motionbench_acc
    aggregation: !function utils.motionbench_aggregate_acc
    higher_is_better: true
  - metric: motionbench_answered_rate
    aggregation: !function utils.motionbench_aggregate_answered_rate
    higher_is_better: true
  - metric: motionbench_tracking
    aggregation: !function utils.motionbench_aggregate_tracking
    higher_is_better: true
lmms_eval_specific_kwargs:
  default:
    pre_prompt: ""
    post_prompt: "Answer with only the option letter (A, B, C, D, E, or F)."
metadata:
  version: 0.1
  description: "MotionBench full integration with per-category tracking and chat/simple compatibility"
