Evals API MVP

[yanxi0830]

DevX Flow

Step 1. Register Eval Dataset

python -m llama_stack.apis.datasets.client

Step 2. Run Eval Scorer

python -m llama_stack.apis.evals.client

• (benchmark) run full preprocess->generation->postprocess->score eval task flow

• (evaluate score only) run scorer only on prepared eval dataset w/ columns expected_answer and generated_answer

LLM As Judge

• Scorer for braintrust AnswerCorrectness()

• Scorer for using judge from LlamaStack distribution hosted models (via inference_api)

Eval Dataset Template Schema

@json_schema_type
class PostprocessedGeneration(BaseModel):
    completion_message: str
    logprobs: Optional[List[TokenLogProbs]] = None

@json_schema_type
class ScorerInputSample(DatasetSample):
    """
    A dataset is required to have the following columns to be used for scoring:
    - generated_answer: str
    - expected_answer: Union[str, List[str]]
    - (optional) input_query: str
    - (optional) generation_output: PostprocessedGeneration
    """

    generated_answer: str
    expected_answer: Union[str, List[str]]
    input_query: Optional[str] = None
    generation_output: Optional[PostprocessedGeneration] = None

High-Level Changes

1. New endpoints API added

• /datasets for registering/deleting datasets

  • /datasets/create --> add new datasets to distribution. supports custom file upload / url / huggingface datasets.
  • /datasets/get
  • /datasets/delete
  • /datasets/list

• /evals for running evaluation tasks

  • /evals/run_eval_task --> run full eval including preprocessing -> generation -> postprocessing -> scoring
  • /evals/run_scorer --> run scoring only

2. New datastructures added

• Registry: for maintaining datasets, scorers, processors
  • BaseScorer: evaluation methods for scoring
  • BaseDataset: supports custom datasets / huggingface datasets
  • BaseGeneratorProcessor: performing preprocessing/postprocessing before/after generation
  • BaseGenerator: performing inference generation

(experimental) eleuther harness

• integrate w/ Eleuther Eval Harness
• add custom task for eleuther eval harness on mmlu_pro (e.g. recipe)
• dummy loglikelihood outputs for inference
  

-- Not in this PR
🚧 jobs API for background eval job scheduling
🚧 batch inference
🚧 persist intermediate datasets during run_eval_task