Upload regenerated NanoRuMTEB

README.md

@@ -0,0 +1,108 @@
+---
+{
+  "configs": [
+    {
+      "config_name": "corpus",
+      "data_files": [
+        {
+          "split": "miracl_ru",
+          "path": "corpus/miracl_ru-00000-of-00001.parquet"
+        },
+        {
+          "split": "ria_news",
+          "path": "corpus/ria_news-00000-of-00001.parquet"
+        },
+        {
+          "split": "ru_bq",
+          "path": "corpus/ru_bq-00000-of-00001.parquet"
+        }
+      ]
+    },
+    {
+      "config_name": "queries",
+      "data_files": [
+        {
+          "split": "miracl_ru",
+          "path": "queries/miracl_ru-00000-of-00001.parquet"
+        },
+        {
+          "split": "ria_news",
+          "path": "queries/ria_news-00000-of-00001.parquet"
+        },
+        {
+          "split": "ru_bq",
+          "path": "queries/ru_bq-00000-of-00001.parquet"
+        }
+      ],
+      "default": true
+    },
+    {
+      "config_name": "qrels",
+      "data_files": [
+        {
+          "split": "miracl_ru",
+          "path": "qrels/miracl_ru-00000-of-00001.parquet"
+        },
+        {
+          "split": "ria_news",
+          "path": "qrels/ria_news-00000-of-00001.parquet"
+        },
+        {
+          "split": "ru_bq",
+          "path": "qrels/ru_bq-00000-of-00001.parquet"
+        }
+      ]
+    },
+    {
+      "config_name": "bm25",
+      "data_files": [
+        {
+          "split": "miracl_ru",
+          "path": "bm25/miracl_ru-00000-of-00001.parquet"
+        },
+        {
+          "split": "ria_news",
+          "path": "bm25/ria_news-00000-of-00001.parquet"
+        },
+        {
+          "split": "ru_bq",
+          "path": "bm25/ru_bq-00000-of-00001.parquet"
+        }
+      ]
+    }
+  ],
+  "tags": [
+    "information-retrieval",
+    "retrieval",
+    "nano",
+    "bm25",
+    "hakari-bench"
+  ]
+}
+---
+
+# NanoRuMTEB
+
+This is a private regenerated Nano-style retrieval dataset for HAKARI-Bench.
+
+## Source Links
+
+- [mteb/MIRACLRetrievalHardNegatives](https://huggingface.co/datasets/mteb/MIRACLRetrievalHardNegatives)
+- [mteb/RiaNewsRetrieval_test_top_250_only_w_correct-v2](https://huggingface.co/datasets/mteb/RiaNewsRetrieval_test_top_250_only_w_correct-v2)
+- [ai-forever/rubq-retrieval](https://huggingface.co/datasets/ai-forever/rubq-retrieval)
+
+## Data Layout
+
+The dataset uses four Hugging Face Datasets configs: `corpus`, `queries`, `qrels`, and `bm25`. Each config contains the same Hub split names. Hub split names replace unsupported characters in the original Nano split names with underscores; the original split names are listed below and stored in `metadata/{hub_split}.json`.
+
+## Split Mapping
+
+| Hub split | Original Nano split | Source task | Source dataset | Queries | Corpus | Qrels |
+|---|---|---|---|---:|---:|---:|
+| miracl_ru | miracl-ru | MIRACLRetrievalHardNegatives.v2 | [mteb/MIRACLRetrievalHardNegatives](https://huggingface.co/datasets/mteb/MIRACLRetrievalHardNegatives) | 200 | 10000 | 579 |
+| ria_news | ria-news | RiaNewsRetrievalHardNegatives.v2 | [mteb/RiaNewsRetrieval_test_top_250_only_w_correct-v2](https://huggingface.co/datasets/mteb/RiaNewsRetrieval_test_top_250_only_w_correct-v2) | 200 | 10000 | 200 |
+| ru_bq | ru-bq | RuBQRetrieval | [ai-forever/rubq-retrieval](https://huggingface.co/datasets/ai-forever/rubq-retrieval) | 200 | 10000 | 334 |
+
+## License
+
+NanoRuMTEB is a derived dataset. Users must comply with the licenses, terms, and attribution requirements of the upstream MTEB task sources and their original datasets.

audit.json

@@ -0,0 +1,54 @@
+{
+  "dataset_dir": "output/nano_datasets/NanoRuMTEB",
+  "subsets": [
+    "ai-forever__rubq-retrieval",
+    "mteb"
+  ],
+  "split_count": 3,
+  "missing_positive_doc_count_after_forcing": 0,
+  "issues": [],
+  "subset_audits": {
+    "ai-forever__rubq-retrieval": {
+      "dataset_dir": "output/nano_datasets/NanoRuMTEB/ai-forever__rubq-retrieval",
+      "splits": [
+        "ru-bq"
+      ],
+      "counts": [
+        {
+          "split_name": "ru-bq",
+          "queries": 200,
+          "corpus": 10000,
+          "qrels": 334,
+          "bm25": 200
+        }
+      ],
+      "missing_positive_doc_count_after_forcing": 0,
+      "issues": []
+    },
+    "mteb": {
+      "dataset_dir": "output/nano_datasets/NanoRuMTEB/mteb",
+      "splits": [
+        "miracl-ru",
+        "ria-news"
+      ],
+      "counts": [
+        {
+          "split_name": "miracl-ru",
+          "queries": 200,
+          "corpus": 10000,
+          "qrels": 579,
+          "bm25": 200
+        },
+        {
+          "split_name": "ria-news",
+          "queries": 200,
+          "corpus": 10000,
+          "qrels": 200,
+          "bm25": 200
+        }
+      ],
+      "missing_positive_doc_count_after_forcing": 0,
+      "issues": []
+    }
+  }
+}

bm25/miracl_ru-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a115f22c15316b51aa08e54f03a0eea68b68b4e3093879f65a7c3c1879c758d
+size 264268

bm25/ria_news-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:435f4a1ac99f6da36397b91af363a8b046b61a4f2728ffdfde1fb02aef24f39e
+size 178334

bm25/ru_bq-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb37b763159298f4792e469cfc058777f3a1b00903978a999a37b93c92abe04b
+size 162601

corpus/miracl_ru-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69fc9afebb63c549142c263d9769d3fd425818af61683134e3b4ebc81fcc9825
+size 4877192

corpus/ria_news-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:529276dd8e1e71a5843c849cf4798ca5285387511c48bc32fb65136667f066bc
+size 10150855

corpus/ru_bq-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd5b62b370a1a8c064732e901e0cf3c5ef54afcf2d371932990f897792f73c4f
+size 4522753

metadata/miracl_ru.json

@@ -0,0 +1,59 @@
+{
+  "split_name": "miracl-ru",
+  "queries": 200,
+  "corpus": 10000,
+  "qrels": 579,
+  "source_positive_qrels": 2833,
+  "source_non_positive_qrels": 7637,
+  "query_limit": 200,
+  "doc_limit": 10000,
+  "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": 577,
+    "source_hard_negative_doc_candidates": 1513,
+    "kept_hard_negative_docs": 1142,
+    "missing_positive_docs": 0,
+    "missing_hard_negative_docs": 369,
+    "duplicate_doc_texts_removed": 62,
+    "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": 0.45902945963760994,
+    "forced_queries": 56,
+    "forced_doc_count": 87,
+    "missing_positive_doc_count_after_forcing": 0
+  },
+  "source_task": "MIRACLRetrievalHardNegatives.v2",
+  "source_dataset_id": "mteb/MIRACLRetrievalHardNegatives",
+  "source_dataset_revision": "d7d94fa4b946cec4a27c84653aa0cf6b33f74a3c",
+  "source_dataset_subset": "ru",
+  "source_eval_split": "dev",
+  "original_nano_split": "miracl-ru",
+  "hub_split": "miracl_ru"
+}

metadata/ria_news.json

@@ -0,0 +1,59 @@
+{
+  "split_name": "ria-news",
+  "queries": 200,
+  "corpus": 10000,
+  "qrels": 200,
+  "source_positive_qrels": 1000,
+  "source_non_positive_qrels": 0,
+  "query_limit": 200,
+  "doc_limit": 10000,
+  "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": 200,
+    "source_hard_negative_doc_candidates": 0,
+    "kept_hard_negative_docs": 0,
+    "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": 0.8479284860934532,
+    "forced_queries": 9,
+    "forced_doc_count": 9,
+    "missing_positive_doc_count_after_forcing": 0
+  },
+  "source_task": "RiaNewsRetrievalHardNegatives.v2",
+  "source_dataset_id": "mteb/RiaNewsRetrieval_test_top_250_only_w_correct-v2",
+  "source_dataset_revision": "d42860a6c15f0a2c4485bda10c6e5b641fdfe479",
+  "source_dataset_subset": "default",
+  "source_eval_split": "test",
+  "original_nano_split": "ria-news",
+  "hub_split": "ria_news"
+}

metadata/ru_bq.json

@@ -0,0 +1,59 @@
+{
+  "split_name": "ru-bq",
+  "queries": 200,
+  "corpus": 10000,
+  "qrels": 334,
+  "source_positive_qrels": 2845,
+  "source_non_positive_qrels": 0,
+  "query_limit": 200,
+  "doc_limit": 10000,
+  "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": 332,
+    "source_hard_negative_doc_candidates": 0,
+    "kept_hard_negative_docs": 0,
+    "missing_positive_docs": 0,
+    "missing_hard_negative_docs": 0,
+    "duplicate_doc_texts_removed": 8,
+    "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": 0.5131650331919719,
+    "forced_queries": 67,
+    "forced_doc_count": 85,
+    "missing_positive_doc_count_after_forcing": 0
+  },
+  "source_task": "RuBQRetrieval",
+  "source_dataset_id": "ai-forever/rubq-retrieval",
+  "source_dataset_revision": "e19b6ffa60b3bc248e0b41f4cc37c26a55c2a67b",
+  "source_dataset_subset": "default",
+  "source_eval_split": "test",
+  "original_nano_split": "ru-bq",
+  "hub_split": "ru_bq"
+}

qrels/miracl_ru-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:066add71d181bcda34302cc192a8e60cee2db52c4fba8fe1f0924afee481e0db
+size 8004

qrels/ria_news-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1286fe52f6c9f11b4fb5095072ad858cf8d81e9981f84664ebe2f8b20ec4e1b
+size 4028

qrels/ru_bq-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fcb9990612a9f3b1290b2ff5c0152f12a2d51b55b085bdd143d882c84a301f25
+size 4455

queries/miracl_ru-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e36894941989e7303b888f2f4c9a96249c5cebe1c6c7d855f6ab982b10942158
+size 12677

queries/ria_news-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2bdafcb9c56c5abb20d1437966b5b06c5a66be4638d71410a1ceac297c35881
+size 16475

queries/ru_bq-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1ac1c83b362a81cc7a872c55bf28a6d1fcc8108f93d32778f07066812e08c13
+size 13488