dataset_path: txiong23/Multi-Crit
dataset_name: default

output_type: generate_until
doc_to_visual: !function utils.multi_crit_doc_to_visual
doc_to_text: !function utils.multi_crit_doc_to_text
doc_to_messages: !function utils.multi_crit_doc_to_messages
doc_to_target: preference
process_results: !function utils.multi_crit_process_results

generation_kwargs:
  # Prevent TaskConfig from injecting its default double-newline stop sequence.
  until: []
  max_new_tokens: 8192
  temperature: 0.6
  top_p: 0.95
  num_beams: 1
  do_sample: true

# Preserve the complete generated judgement for the official last-line parser.
# This controls evaluator-side tag removal; it does not disable model reasoning.
reasoning_tags: none

metadata:
  version: 1.0
