sample-config-t5.yaml

# defaults are from llm_evaluate/configs/*.yaml
defaults:
- configs@benchmarks: benchmarks
- configs@datasets: datasets
- configs@extensions: extensions
- configs@metrics: metrics
- configs@model: model
- configs@model_loaders: model_loaders
- configs@parsers: parsers
- _self_


datasets:
  demo-chat:
    tasks:
      random-questions:
        task_type: generation
    column_input: prompt
    column_reference: response
    description: Dataset for random generative tasks (single)
    metadata:
      source: local
      data:
      - prompt: How to get to Vancouver from Seattle?
        response: Take a plane, bus, train or car.
      - prompt: Generate 3 male names in a list.
        response: One, Two, Three
      - prompt: Convert 4.33 x 10 to scientific notation.
        response: 4.33e2

  demo-classification:
    tasks:
      random-mcq:
        task_type: classification
        model_output_parser: KeywordParser
        labels:
        - A
        - B
        - C
        - D
        - E
        none_value: E
    column_input: prompt
    column_reference: response
    description: Dataset for random generative tasks (numeric reference)
    metadata:
      source: local
      data:
      - prompt: |
          Pick the correct answer for the following question:
          1 + 1 = ?
          A. 0
          B. 1
          C. 2
          D. 3
        response: C
      - prompt: |
          Answer the following multiple choice question:
          2 + 2 = ?
          A. 1
          B. 2
          C. 3
          D. 4
        response: D
      - prompt: |
          Select the correct response for the following multiple choice question:
          1 + 2 = ?
          A. 1
          B. 2
          C. 3
          D. 4
        response: C


benchmarks:
  demo:
    demo-chat:
      tasks:
        random-questions:
          metrics:
          - rouge
      prompt: Answer the following question as concisely as possible.

    demo-classification:
      tasks:
        random-mcq:
          metrics:
          - accuracy
          - precision
          - f1
          - recall


model:
  model: google/t5-efficient-tiny
  model_type: hf-automodel
  tokenizer_args:
    model_max_length: 3000
    truncation_side: right
    truncation: longest_first
  model_load_args:
    task_type: T5ForConditionalGeneration
    device_map: cpu
    max_input_tokens: 3000
    torch_dtype: auto
  model_inference_args:
    max_new_tokens: 512
    num_beams: 2
    do_sample: False
  add_to_prompt_start: '[Prompt]'
  add_to_prompt_end: '[Response]'