{
  "split_name": "fa",
  "queries": 200,
  "corpus": 10000,
  "qrels": 427,
  "source_positive_qrels": 427,
  "source_non_positive_qrels": 1664,
  "query_limit": 200,
  "doc_limit": 10000,
  "dedupe_query_texts": true,
  "dedupe_doc_texts": true,
  "qrels_score_policy": "score > 0 kept as qrels; score <= 0 excluded and used as hard-negative corpus candidates",
  "query_selection": {
    "skipped_queries_without_positive_qrels": 0,
    "skipped_duplicate_query_ids": 0,
    "skipped_duplicate_query_texts": 0
  },
  "corpus_selection": {
    "positive_doc_count": 426,
    "source_hard_negative_doc_candidates": 1656,
    "kept_hard_negative_docs": 1656,
    "missing_positive_docs": 0,
    "missing_hard_negative_docs": 0,
    "duplicate_doc_texts_removed": 0,
    "hard_negative_sampling_policy": "query_round_robin"
  },
  "qrels_selection": {
    "qrels_per_query_cap": 100,
    "removed_qrels_over_cap": 0,
    "qrels_cap_policy": "positive qrels are capped per query to the BM25 candidate top_k"
  },
  "bm25": {
    "config": {
      "backend": "bm25s",
      "algorithm": "okapi",
      "source": "computed_bm25s",
      "tokenizer": "regex",
      "tokenizer_name": null,
      "stemmer_algorithm": "english",
      "top_k": 100,
      "k1": 1.5,
      "b": 0.75,
      "show_progress": false,
      "auto_selected": false,
      "auto_detected_language": null,
      "auto_detection_language_counts": null,
      "auto_detection_sample_size": null
    },
    "ndcg_at_10": 0.5787921084448737,
    "candidate_coverage": {
      "top_k": 100,
      "query_count": 200,
      "query_with_relevance_count": 200,
      "covered_query_count": 200,
      "query_coverage": 1.0,
      "relevant_count": 427,
      "covered_relevant_count": 427,
      "relevant_coverage": 1.0
    },
    "forced_queries": 14,
    "forced_doc_count": 17,
    "missing_positive_doc_count_after_forcing": 0
  },
  "description": "NanoMIRACL rebuilt from hotchpotch/miracl-hf-unified dev queries, preserving all source positive passages for each sampled query and using source negatives as hard negative corpus candidates before random corpus fill.",
  "source_dataset_id": "hotchpotch/miracl-hf-unified",
  "source_dataset_revision": "21ad00eb467639e927b5badb7c49f4947c6c24ca",
  "source_dataset_subset": "fa_queries",
  "source_eval_split": "dev",
  "source_split_policy": "Use `dev` from `fa_queries` for each language.",
  "source_query_selection_policy": "Take the first 200 queries with at least one source positive; fewer if the source split has fewer eligible queries."
}