dataset_path: FlagEval/EmbSpatial-Bench
test_split: test
output_type: generate_until
doc_to_visual: !function utils.embspatial_doc_to_visual
doc_to_text: !function utils.embspatial_doc_to_text
doc_to_messages: !function utils.embspatial_doc_to_messages
doc_to_target: "answer"

process_results: !function utils.embspatial_process_results

metric_list:
  - metric: embspatial_acc
    aggregation: !function utils.embspatial_aggregate_results
    higher_is_better: true
  - metric: ai2thor_accuracy
    aggregation: !function utils.embspatial_aggregate_ai2thor_accuracy
    higher_is_better: true
  - metric: mp3d_accuracy
    aggregation: !function utils.embspatial_aggregate_mp3d_accuracy
    higher_is_better: true
  - metric: scannet_accuracy
    aggregation: !function utils.embspatial_aggregate_scannet_accuracy
    higher_is_better: true

lmms_eval_specific_kwargs:
  default:
    pre_prompt: "Note: You only need to respond with A, B, C, D without providing any additional information.\n"
    post_prompt: ""
  neo_ov:
    pre_prompt: ""
    post_prompt: ""
    prompt_format: neo_ov

model_specific_generation_kwargs:
  neo_ov:
    patch_size: 16
    min_pixels: 691200
    max_pixels: 691200
    downsample_ratio: 0.5
    system_prompt: ""
    max_new_tokens: 4096
    temperature: null
    top_p: null
    repetition_penalty: null

generation_kwargs:
  max_new_tokens: 1024
