Update dataset README for reranking_hybrid candidates

README.md

@@ -145,161 +145,171 @@ dataset_info:
   download_size: 807492
   dataset_size: 803136
 configs:
-- config_name: bm25
+- config_name: corpus
   data_files:
   - split: NanoBIRCOArguAna
-    path: bm25/NanoBIRCOArguAna-*
+    path: corpus/NanoBIRCOArguAna.parquet
   - split: NanoBIRCOClinicalTrial
-    path: bm25/NanoBIRCOClinicalTrial-*
+    path: corpus/NanoBIRCOClinicalTrial.parquet
   - split: NanoBIRCODorisMae
-    path: bm25/NanoBIRCODorisMae-*
+    path: corpus/NanoBIRCODorisMae.parquet
   - split: NanoBIRCORelic
-    path: bm25/NanoBIRCORelic-*
+    path: corpus/NanoBIRCORelic.parquet
   - split: NanoBIRCOWTB
-    path: bm25/NanoBIRCOWTB-*
-- config_name: corpus
+    path: corpus/NanoBIRCOWTB.parquet
+- config_name: queries
   data_files:
   - split: NanoBIRCOArguAna
-    path: corpus/NanoBIRCOArguAna-*
+    path: queries/NanoBIRCOArguAna.parquet
   - split: NanoBIRCOClinicalTrial
-    path: corpus/NanoBIRCOClinicalTrial-*
+    path: queries/NanoBIRCOClinicalTrial.parquet
   - split: NanoBIRCODorisMae
-    path: corpus/NanoBIRCODorisMae-*
+    path: queries/NanoBIRCODorisMae.parquet
   - split: NanoBIRCORelic
-    path: corpus/NanoBIRCORelic-*
+    path: queries/NanoBIRCORelic.parquet
   - split: NanoBIRCOWTB
-    path: corpus/NanoBIRCOWTB-*
-- config_name: harrier_oss_v1_270m
+    path: queries/NanoBIRCOWTB.parquet
+  default: true
+- config_name: qrels
   data_files:
   - split: NanoBIRCOArguAna
-    path: harrier_oss_v1_270m/NanoBIRCOArguAna-*
+    path: qrels/NanoBIRCOArguAna.parquet
   - split: NanoBIRCOClinicalTrial
-    path: harrier_oss_v1_270m/NanoBIRCOClinicalTrial-*
+    path: qrels/NanoBIRCOClinicalTrial.parquet
   - split: NanoBIRCODorisMae
-    path: harrier_oss_v1_270m/NanoBIRCODorisMae-*
+    path: qrels/NanoBIRCODorisMae.parquet
   - split: NanoBIRCORelic
-    path: harrier_oss_v1_270m/NanoBIRCORelic-*
+    path: qrels/NanoBIRCORelic.parquet
   - split: NanoBIRCOWTB
-    path: harrier_oss_v1_270m/NanoBIRCOWTB-*
-- config_name: qrels
+    path: qrels/NanoBIRCOWTB.parquet
+- config_name: bm25
   data_files:
   - split: NanoBIRCOArguAna
-    path: qrels/NanoBIRCOArguAna-*
+    path: bm25/NanoBIRCOArguAna.parquet
   - split: NanoBIRCOClinicalTrial
-    path: qrels/NanoBIRCOClinicalTrial-*
+    path: bm25/NanoBIRCOClinicalTrial.parquet
   - split: NanoBIRCODorisMae
-    path: qrels/NanoBIRCODorisMae-*
+    path: bm25/NanoBIRCODorisMae.parquet
   - split: NanoBIRCORelic
-    path: qrels/NanoBIRCORelic-*
+    path: bm25/NanoBIRCORelic.parquet
   - split: NanoBIRCOWTB
-    path: qrels/NanoBIRCOWTB-*
-- config_name: queries
+    path: bm25/NanoBIRCOWTB.parquet
+- config_name: harrier_oss_v1_270m
   data_files:
   - split: NanoBIRCOArguAna
-    path: queries/NanoBIRCOArguAna-*
+    path: harrier_oss_v1_270m/NanoBIRCOArguAna.parquet
   - split: NanoBIRCOClinicalTrial
-    path: queries/NanoBIRCOClinicalTrial-*
+    path: harrier_oss_v1_270m/NanoBIRCOClinicalTrial.parquet
   - split: NanoBIRCODorisMae
-    path: queries/NanoBIRCODorisMae-*
+    path: harrier_oss_v1_270m/NanoBIRCODorisMae.parquet
   - split: NanoBIRCORelic
-    path: queries/NanoBIRCORelic-*
+    path: harrier_oss_v1_270m/NanoBIRCORelic.parquet
   - split: NanoBIRCOWTB
-    path: queries/NanoBIRCOWTB-*
-  default: true
+    path: harrier_oss_v1_270m/NanoBIRCOWTB.parquet
 - config_name: reranking_hybrid
   data_files:
   - split: NanoBIRCOArguAna
-    path: reranking_hybrid/NanoBIRCOArguAna-*
+    path: reranking_hybrid/NanoBIRCOArguAna.parquet
   - split: NanoBIRCOClinicalTrial
-    path: reranking_hybrid/NanoBIRCOClinicalTrial-*
+    path: reranking_hybrid/NanoBIRCOClinicalTrial.parquet
   - split: NanoBIRCODorisMae
-    path: reranking_hybrid/NanoBIRCODorisMae-*
+    path: reranking_hybrid/NanoBIRCODorisMae.parquet
   - split: NanoBIRCORelic
-    path: reranking_hybrid/NanoBIRCORelic-*
+    path: reranking_hybrid/NanoBIRCORelic.parquet
   - split: NanoBIRCOWTB
-    path: reranking_hybrid/NanoBIRCOWTB-*
+    path: reranking_hybrid/NanoBIRCOWTB.parquet
+tags:
+- information-retrieval
+- retrieval
+- nano
+- bm25
+- dense-retrieval
+- reranking
+- hakari-bench
 ---
-
 # NanoBIRCO
 
-This dataset is a Nano-style retrieval dataset. Nano-series evaluation can be run easily with [HAKARI-Bench](https://github.com/hotchpotch/hakari-bench).
-
-NanoBIRCO is derived from BIRCO. It follows the Hugging Face Datasets layout convention used by [sentence-transformers/NanoBEIR-en](https://huggingface.co/datasets/sentence-transformers/NanoBEIR-en): each Nano split has separate `corpus`, `queries`, and `qrels` tables, and BM25 candidates are provided separately in a `bm25` table. This layout follows the NanoBEIR-style evaluation approach summarized in [NanoBEIR](https://huggingface.co/blog/sionic-ai/eval-sionic-nano-beir).
+This dataset is a Nano-style retrieval dataset for [HAKARI-bench](https://github.com/hakari-bench/hakari-bench).
 
 NanoBIRCO contains 5 Nano retrieval splits derived from BIRCO. Each split keeps up to 200 eligible queries and up to 10000 corpus documents, with exact duplicate query and document text removed where the generator records that policy.
 
-## Source Links
+## Usage
 
-- Source benchmark: `BIRCO`
-- Source benchmark repository: https://github.com/embeddings-benchmark/mteb
-- `mteb/BIRCO-ArguAna-Test`: https://huggingface.co/datasets/mteb/BIRCO-ArguAna-Test
-- `mteb/BIRCO-ClinicalTrial-Test`: https://huggingface.co/datasets/mteb/BIRCO-ClinicalTrial-Test
-- `mteb/BIRCO-DorisMae-Test`: https://huggingface.co/datasets/mteb/BIRCO-DorisMae-Test
-- `mteb/BIRCO-Relic-Test`: https://huggingface.co/datasets/mteb/BIRCO-Relic-Test
-- `mteb/BIRCO-WTB-Test`: https://huggingface.co/datasets/mteb/BIRCO-WTB-Test
+```python
+from datasets import load_dataset
+
+dataset_id = "hakari-bench/NanoBIRCO"
+split = "NanoBIRCOArguAna"
+
+queries = load_dataset(dataset_id, "queries", split=split)
+corpus = load_dataset(dataset_id, "corpus", split=split)
+qrels = load_dataset(dataset_id, "qrels", split=split)
+reranking_candidates = load_dataset(dataset_id, "reranking_hybrid", split=split)
+```
 
 ## Data Layout
 
-This dataset uses four Hugging Face Datasets configs:
+This dataset uses six Hugging Face Datasets configs:
 
 - `corpus`: documents with `_id` and `text`
 - `queries`: queries with `_id` and `text`
 - `qrels`: positive relevance labels with `query-id` and `corpus-id`
 - `bm25`: BM25 candidate lists with `query-id` and `corpus-ids`
+- `harrier_oss_v1_270m`: dense candidate lists from `microsoft/harrier-oss-v1-270m`
+- `reranking_hybrid`: RRF candidate lists built from `bm25` and `harrier_oss_v1_270m`
 
-Each config uses the same Nano split names. If the actual generated dataset uses a different schema, config name, path layout, or field name, revise this section before publishing the README.
-
-## Construction Steps
+Each config has the same Nano split names.
 
-This dataset was built as follows. If the actual generation procedure differs, revise this section before publishing the README.
+## Candidate Construction
 
-1. Use BIRCO as the upstream benchmark or dataset family.
-2. Load the source datasets recorded in `manifest.json` and per-split metadata files.
-3. Use the evaluation splits selected by `scripts/create_mteb_nano_series_local.py` as the source evaluation split policy.
-4. Create one Nano split for each selected source retrieval task.
-5. Keep up to 200 eligible queries per Nano split.
-6. Include qrels-positive documents for the selected queries. For graded qrels, only rows with `score > 0` are treated as positive.
-7. Fill the corpus from source corpus order up to 10000 documents.
-8. Remove exact duplicate document text within each split. If a removed duplicate was referenced by qrels, rewrite qrels to the kept document id when the generator records that policy.
-9. Store document title and body as a single `text` field when the source provides both.
-10. Generate BM25 top-100 candidates with the tokenization policy recorded per split.
-11. If a qrels-positive document is missing from the raw BM25 result, insert it into the final `bm25` candidate list by replacing a tail non-positive candidate.
+- `bm25`: local BM25 top-500 with automatic language-aware tokenization. The resolved tokenizer is shown in the Candidate Quality table, for example `wordseg@ja`.
+- `harrier_oss_v1_270m`: dense top-500 from `microsoft/harrier-oss-v1-270m`. In tables this is shown as `Dense`; Dense means `microsoft/harrier-oss-v1-270m` with the `web_search_query` prompt for queries and cosine similarity over normalized embeddings.
+- `reranking_hybrid`: RRF over `bm25` and `harrier_oss_v1_270m` using `rrf_k=100`, keeping the RRF top-100.
 
-Some source queries had more positives than the BM25 candidate cap. For those cases, the Nano qrels keep the first available positive document so BM25 top-k remains comparable; 44 selected queries were capped and 4851 source positive qrels were omitted from this Nano slice.
+Safeguard means rank 101 is appended only when RRF top-100 contains no qrels-positive document.
 
-## BM25 Subset Policy
+## Split Statistics
 
-The `bm25` config is a candidate subset for first-stage retrieval and reranking. It is not a separate source dataset. Each row contains one query id and a ranked list of corpus ids.
+Length statistics are character counts computed with `len(str(text))`.
 
-BM25 candidates are generated from the selected corpus for each split. The configured candidate cap is top-100. When a qrels-positive document is not present in the raw BM25 result, the missing positive is forced into the final candidate list by replacing a tail candidate that is not positive for that query. Candidate ids are kept unique after replacement.
+| Nano split | Queries | Corpus | Qrels | Query chars avg | Query chars p50 | Query chars p75 | Doc chars avg | Doc chars p50 | Doc chars p75 |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| NanoBIRCOArguAna | 98 | 3081 | 98 | 1124.0 | 1101.5 | 1410.2 | 1140.1 | 1082.0 | 1445.0 |
+| NanoBIRCOClinicalTrial | 50 | 3375 | 1042 | 497.0 | 438.5 | 558.8 | 1174.3 | 1421.0 | 1579.0 |
+| NanoBIRCODorisMae | 60 | 5544 | 1569 | 995.5 | 993.5 | 1095.5 | 1220.3 | 1208.5 | 1431.0 |
+| NanoBIRCORelic | 100 | 5023 | 100 | 1016.3 | 1054.0 | 1144.5 | 477.3 | 438.0 | 627.0 |
+| NanoBIRCOWTB | 100 | 1766 | 100 | 811.3 | 788.5 | 954.8 | 1091.2 | 1108.0 | 1269.8 |
 
-## Split Mapping
+## Candidate Quality
 
-| Nano split | Source task | Source dataset | Queries | Corpus | Qrels |
-|---|---|---|---:|---:|---:|
-| `NanoBIRCOArguAna` | `BIRCO-ArguAna` | `mteb/BIRCO-ArguAna-Test` | 98 | 3081 | 98 |
-| `NanoBIRCOClinicalTrial` | `BIRCO-ClinicalTrial` | `mteb/BIRCO-ClinicalTrial-Test` | 50 | 3375 | 1042 |
-| `NanoBIRCODorisMae` | `BIRCO-DorisMae` | `mteb/BIRCO-DorisMae-Test` | 60 | 5544 | 1569 |
-| `NanoBIRCORelic` | `BIRCO-Relic` | `mteb/BIRCO-Relic-Test` | 100 | 5023 | 100 |
-| `NanoBIRCOWTB` | `BIRCO-WTB` | `mteb/BIRCO-WTB-Test` | 100 | 1766 | 100 |
+`nDCG@10` and `Recall@100` are computed from the included candidate rankings against the included qrels, then reported as 0-100 scores such as `52.45`. `Recall@100` uses only the top 100 candidates; an optional rank-101 safeguard positive is not counted in `Recall@100`.
 
-## BM25 nDCG@10
+Dense means `microsoft/harrier-oss-v1-270m` with the `web_search_query` prompt and cosine similarity.
 
-`nDCG@10` is computed from the included BM25 ranking against the included qrels.
+| Nano split | BM25 tokenizer | BM25 nDCG@10 | Dense nDCG@10 | Hybrid nDCG@10 | BM25 Recall@100 | Dense Recall@100 | Hybrid Recall@100 | Hybrid candidates | Safeguard positives |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| Mean | - | 26.93 | 29.59 | 31.11 | 63.53 | 70.96 | 73.38 | - | 66 |
+| NanoBIRCOArguAna | english_porter_stop | 42.93 | 50.62 | 49.32 | 97.96 | 97.96 | 100.00 | 100 | 0 |
+| NanoBIRCOClinicalTrial | english_porter_stop | 13.22 | 21.52 | 19.59 | 31.03 | 48.52 | 45.61 | 100-101 | 1 |
+| NanoBIRCODorisMae | english_porter_stop | 38.66 | 41.40 | 40.12 | 70.68 | 72.34 | 80.30 | 100-101 | 6 |
+| NanoBIRCORelic | english_porter_stop | 13.14 | 7.25 | 12.76 | 59.00 | 66.00 | 69.00 | 100-101 | 31 |
+| NanoBIRCOWTB | english_porter_stop | 26.69 | 27.14 | 33.76 | 59.00 | 70.00 | 72.00 | 100-101 | 28 |
 
-Tokenizer policy summary: `stemmer:en`.
+## Hybrid Safeguard Summary
 
-| Nano split | Tokenizer | Forced BM25 positives | BM25 nDCG@10 |
-|---|---|---:|---:|
-| `NanoBIRCOArguAna` | `stemmer:en` | 6 | 0.4051 |
-| `NanoBIRCOClinicalTrial` | `stemmer:en` | 819 | 0.1194 |
-| `NanoBIRCODorisMae` | `stemmer:en` | 942 | 0.2469 |
-| `NanoBIRCORelic` | `stemmer:en` | 62 | 0.0633 |
-| `NanoBIRCOWTB` | `stemmer:en` | 65 | 0.0751 |
+- Safeguard positives: 66
+- Rows limited by corpus size: 0
+- Metadata file: `reranking_hybrid_metadata.json`
 
-## Skipped Tasks
+## Source Links
 
-No source tasks were skipped.
+- Source benchmark: `BIRCO`
+- Source benchmark repository: https://github.com/embeddings-benchmark/mteb
+- `mteb/BIRCO-ArguAna-Test`: https://huggingface.co/datasets/mteb/BIRCO-ArguAna-Test
+- `mteb/BIRCO-ClinicalTrial-Test`: https://huggingface.co/datasets/mteb/BIRCO-ClinicalTrial-Test
+- `mteb/BIRCO-DorisMae-Test`: https://huggingface.co/datasets/mteb/BIRCO-DorisMae-Test
+- `mteb/BIRCO-Relic-Test`: https://huggingface.co/datasets/mteb/BIRCO-Relic-Test
+- `mteb/BIRCO-WTB-Test`: https://huggingface.co/datasets/mteb/BIRCO-WTB-Test
 
 ## License