def run_multi_hop()

in yourbench/pipeline/question_generation.py [0:0]

• 41 lines of code
• 11 McCabe index (conditional complexity)

def run_multi_hop(config: dict[str, Any]) -> None:
    """
    Orchestrates both multi-hop and cross-document question generation pipelines,
    if enabled in config
    """
    stage_cfg = config.get("pipeline", {}).get(MULTI_HOP_KEY, {})
    cross_cfg = stage_cfg.get("cross_document", {})
    run_multi = stage_cfg.get("run", False)
    run_cross = cross_cfg.get("enable", False)

    if not run_multi:
        logger.info("Multi-hop question generation is disabled.")
        return

    question_mode = stage_cfg.get("question_mode", "open-ended")
    if question_mode not in {"open-ended", "multi-choice"}:
        logger.warning(f"Invalid question_mode '{question_mode}', defaulting to 'open-ended'")
        question_mode = "open-ended"

    system_prompt = (
        MULTI_HOP_QUESTION_GENERATION_SYSTEM_PROMPT_MULTI
        if question_mode == "multi-choice"
        else MULTI_HOP_QUESTION_GENERATION_SYSTEM_PROMPT
    )
    system_msg = {"role": "system", "content": system_prompt}

    chunked_ds = custom_load_dataset(config=config, subset="chunked")
    logger.info(f"Loaded {len(chunked_ds)} documents for multi-hop processing.")

    def _run_and_save(dataset, label: str):
        if not dataset or len(dataset) == 0:
            logger.warning(f"No valid {label} dataset found. Skipping.")
            return
        inference_calls, inference_index_map = build_multi_hop_inference_calls(dataset, system_msg, stage_cfg)
        if not inference_calls:
            logger.warning(f"No valid inference calls for {label}.")
            return
        responses = run_inference(config=config, step_name=MULTI_HOP_KEY, inference_calls=inference_calls)
        final_rows = parse_multi_hop_responses(responses, inference_index_map, stage_cfg)
        if final_rows:
            logger.info(f"Saving {len(final_rows)} {label} questions.")
            custom_save_dataset(Dataset.from_list(final_rows), config=config, subset=label)
        else:
            logger.info(f"No valid {label} questions parsed.")

    # Run standard multi-hop if enabled
    _run_and_save(chunked_ds, "multi_hop_questions")

    # Run cross-document if enabled
    if run_cross:
        logger.info("Starting cross-document question generation.")
        cross_ds = _create_cross_document_dataset(chunked_ds, cross_cfg)
        logger.info(f"Generated {len(cross_ds)} cross-document combinations.")
        _run_and_save(cross_ds, "cross_document_questions")