Datasets:
run string | created_at string | corpus_chunks int64 | method string | limitation string | summary dict | items list |
|---|---|---|---|---|---|---|
m2-baseline | 2026-10-24T17:21:52.965963+00:00 | 227,449 | subject+15-word queries from held-out chunks; exact chunk/section match | queries share vocabulary with gold chunks, favoring BM25; sanity baseline only | {
"n": 200,
"k": 50,
"chunk_recall@5": {
"dense": 0.505,
"bm25": 0.61,
"rrf": 0.61
},
"chunk_recall@20": {
"dense": 0.78,
"bm25": 0.995,
"rrf": 0.955
},
"section_recall@20": {
"dense": 0.805,
"bm25": 1,
"rrf": 0.975
}
} | [
{
"chunk_id": "25.606_2024_1",
"section": "25.606",
"year": 2024,
"query": "Postaward determinations. ficer may deny the request. (b) The contracting officer must base evaluation of any request",
"dense_r5": 0,
"dense_r20": 0,
"bm25_r5": 1,
"bm25_r20": 1,
"rrf_r5": 0,
"rrf_r2... |
ClauseGym-Bench
Frozen benchmark for version-aware retrieval over FAR/DFARS 2017–2025 (US public domain). Companion to https://github.com/raihan-js/clausegym.
asof_questions.jsonl(700): 500 explicit-date + 200 change questions with gold chunk IDs. Queries use section id + subject only.diffs.jsonl(9,604): consecutive-year changed section pairs with character counts and 500-char excerpts.leakage_ladder.json: naive 0.802 → range 0.522 → graph 0.000 top-1 leakage on the frozen set (rule-based scoring, no LLM judge).eval_retrieval.json: M2 sanity baseline (n=200).asof_meta.json: provenance (created_at, source rows).
Corpus behind the IDs: 227,449 bge-m3 chunks in Postgres+pgvector (rebuild with the repo scripts; embeddings not included).
- Downloads last month
- 32