Add dataset-from-file command to extract datasets from benchmark reports

[Harshith-umesh]

This PR adds a new preprocessing command that enables users to extract datasets from saved benchmark reports, facilitating "apples-to-apples" model comparisons and reproducible benchmarking workflows.

New guidellm preprocess dataset-from-file command that converts benchmark report files into reusable datasets.

Users often want to compare different models using identical prompts to eliminate variability. This command extracts successful request-response pairs from benchmark results, creating standardized datasets with known prompt and output token counts.

The command validates benchmark reports and extracts successful requests, and outputs a structured JSON dataset.

Features Added

• guidellm preprocess dataset-from-file [OPTIONS] BENCHMARK_FILE
• Uses GenerativeBenchmarksReport.load_file() for comprehensive input validation
• User-friendly error messages without Python tracebacks
• Optional --show-stats flag for dataset analysis

Testing

./venv/bin/python -m pytest tests/unit/entrypoints/test_dataset_from_file_entrypoint.py -v

Usage Examples

# Extract dataset from benchmark results
guidellm preprocess dataset-from-file benchmark-results.json -o my_dataset.json

# Extract and Show statistics about the extracted dataset
guidellm preprocess dataset-from-file benchmark-results.json --show-stats