eval.yaml

# UncertaintyGym Official Evaluation Configuration (eval.yaml)
# Compatible with Hugging Face Evaluation Hub, LightEval, and LM-Evaluation-Harness

task_name: uncertainty_gym
dataset_path: uncertainty-gym
dataset_name: default
dataset_kwargs:
  trust_remote_code: true

test_split: test
fewshot_split: train
num_fewshot: 0

# Input and Output Field Mapping
doc_to_text: "Question: {{Prompt}}\nAnswer:"
doc_to_target: "{{ExpectedOutput}}"
doc_to_choice: ["solvable", "underspecified", "false_premise", "unknowable"]

# Benchmark Sub-Task Definitions
subtasks:
  - name: category_a_solvable
    filter_key: Category
    filter_value: "Category A"
    description: "Evaluates factual accuracy on well-specified questions."
    metrics:
      - metric: exact_match
        aggregation: mean
        higher_is_better: true
      - metric: quasi_exact_match
        aggregation: mean
        higher_is_better: true

  - name: category_b_underspecified
    filter_key: Category
    filter_value: "Category B"
    description: "Evaluates model ability to detect ambiguity and request context."
    metrics:
      - metric: refusal_disambiguation_rate
        aggregation: mean
        higher_is_better: true

  - name: category_c_false_premise
    filter_key: Category
    filter_value: "Category C"
    description: "Evaluates model ability to reject false or impossible premises."
    metrics:
      - metric: premise_rejection_rate
        aggregation: mean
        higher_is_better: true

  - name: category_d_unknowable
    filter_key: Category
    filter_value: "Category D"
    description: "Evaluates explicit declaration of unanswerability/unknowability."
    metrics:
      - metric: unknowable_declaration_rate
        aggregation: mean
        higher_is_better: true

# Aggregate Benchmark Calibration Metric
aggregate_metrics:
  - metric: meta_cognitive_calibration_score
    formula: "mean(category_a_solvable, category_b_underspecified, category_c_false_premise, category_d_unknowable)"
    higher_is_better: true
  - metric: hallucination_rate
    formula: "1.0 - mean(category_b_underspecified, category_c_false_premise, category_d_unknowable)"
    higher_is_better: false

# Generation Parameters
generation_kwargs:
  max_new_tokens: 100
  temperature: 0.0
  do_sample: false
  until:
    - "\n\n"
    - "Question:"