dataset_path: json
dataset_kwargs:
  data_files:
    train: hf://datasets/camelCase12/vpct-1/sim_*_results.json
  token: false
task: vpct
test_split: train
output_type: generate_until
process_docs: !function utils.vpct_process_docs
doc_to_visual: !function utils.vpct_doc_to_visual
doc_to_text: !function utils.vpct_doc_to_text
doc_to_target: !function utils.vpct_doc_to_target
process_results: !function utils.vpct_process_results
metric_list:
  - metric: vpct_accuracy
    aggregation: !function utils.vpct_aggregate_accuracy
    higher_is_better: true
  - metric: vpct_answered_rate
    aggregation: !function utils.vpct_aggregate_answered_rate
    higher_is_better: true
generation_kwargs:
  max_new_tokens: 32
  temperature: 0
  do_sample: false
lmms_eval_specific_kwargs:
  default:
    pre_prompt: ""
    post_prompt: "\nRespond with answer(X), where X is 1, 2, or 3."
metadata:
  version: 0.0
