Analyze Dataset Llm

Run LLM-based dataset quality analysis.

This endpoint performs the slow (~8s) LLM-based diversity and quality analysis separately from the fast statistical analysis endpoint.

Use /felix/dataset/analyze for fast statistical metrics (distribution, duplicates, outliers, vendi score, visualization). Use this endpoint when you need the LLM's reasoning about dataset quality and reducibility.

Provide data inline via dataset or reference a stored dataset with dataset_name.

Request body

task_type'ner' | 'classification' | 'generative' required

Task type of the dataset

task_descriptionstring nullable

Description of the task/domain for context. Helps the LLM understand the intended use case.

labelsstring[] nullable

List of label names for the task. If not provided, will be extracted from the dataset.

datasetobject[] nullable

List of data samples (optional if dataset_name provided)

dataset_namestring nullable

Name of stored dataset to analyze (optional if dataset provided)

dataset_versionstring nullable

Dataset version (latest if omitted)

Example request

{
  "dataset_name": "my-ner-dataset",
  "labels": [
    "ORG",
    "LOC",
    "PERSON"
  ],
  "task_description": "Extract company names and locations from news articles",
  "task_type": "ner"
}

Response

Successful Response

reasoning_tracestring required

Step-by-step reasoning trace showing how the model analyzed the dataset's diversity, similar to an ML engineer's thought process

summarystring required

Concise summary of the diversity assessment

diversity_rating'low' | 'moderate' | 'high' | 'excellent' required

Overall diversity rating based on the analysis

key_observationsstring[]

Key observations about the dataset's diversity

recommendationsstring[]

Actionable recommendations to improve diversity if needed

model_usedstring required

The LLM model used for the analysis

v2

Analyze Dataset Llm

Request body

Example request

Response