Upload README.md with huggingface_hub

README.md

@@ -1,50 +1,125 @@
 ---
-dataset_info:
-- config_name: corpus
-  features:
-  - name: document_id
-    dtype: string
-  - name: page_number
-    dtype: int32
-  - name: image
-    dtype: image
-  splits:
-  - name: test
-    num_bytes: 5869749848
-    num_examples: 16247
-  download_size: 5869702871
-  dataset_size: 5869749848
-- config_name: queries
-  features:
-  - name: query
-    dtype: string
-  - name: positive_passages
-    list: int32
-  - name: soft_negative_passages
-    list: int32
-  - name: hard_negative_passages
-    list: int32
-  - name: gold_in_top_100
-    dtype: bool
-  - name: gold_in
-    dtype: int32
-  - name: text_only
-    dtype: bool
-  - name: topics
-    list: string
-  splits:
-  - name: test
-    num_bytes: 3857592
-    num_examples: 10760
-  download_size: 2763068
-  dataset_size: 3857592
+license: cc-by-4.0
+language:
+  - ru
+task_categories:
+  - visual-document-retrieval
+  - text-retrieval
+tags:
+  - colbert
+  - multi-vector
+  - document-retrieval
+  - russian
+  - business-documents
+  - late-interaction
 configs:
-- config_name: corpus
-  data_files:
-  - split: test
-    path: corpus/test-*
-- config_name: queries
-  data_files:
-  - split: test
-    path: queries/test-*
+  - config_name: corpus
+    default: true
+    data_files:
+      - split: test
+        path: "corpus/test-*"
+  - config_name: queries
+    data_files:
+      - split: test
+        path: "queries/test-*"
 ---
+
+# RuDoRe — Russian Document Retrieval (eval split)
+
+The **judge-aligned eval split** of the RuDoRe benchmark for
+Russian-language visual document retrieval. Pages are screenshots of
+business-style Russian documents (presentations, reports, marketing
+decks, course materials). Queries are Russian, generated by a
+vision-language model conditioned on each page and then rephrased to
+remove the lexical-copy artifact that inflates retrieval metrics on
+machine-generated queries. Every sample in this release is
+**judge-aligned** by construction (it survived an LLM-judge filter at
+build time).
+
+This release contains the **test split only**. The train split is not
+distributed.
+
+## Configs
+
+| Config | Split | Description |
+| --- | --- | --- |
+| `corpus` | `test` | One row per page: PNG image plus page / document metadata. |
+| `queries` | `test` | One row per query: query text and gold `positive_passages` index into `corpus`. |
+
+End users load it with:
+
+```python
+from datasets import load_dataset
+
+corpus  = load_dataset("rudore/RuDoRe", "corpus",  split="test")
+queries = load_dataset("rudore/RuDoRe", "queries", split="test")
+```
+
+## Statistics
+
+| Quantity | Count |
+| --- | ---: |
+| Documents | 174 |
+| Pages (corpus rows) | 6,383 |
+| Queries (test) | 18,805 |
+| Topics | 13 |
+
+### Per-topic test-query counts
+
+The eval split is topic-balanced via a per-topic quota cap at build
+time; topics with a smaller crawl end below the cap.
+
+| Topic | Test pairs |
+| --- | ---: |
+| Academic course decks | 3,589 |
+| Computer science | 2,330 |
+| Natural sciences | 2,075 |
+| Medicine & biology | 1,523 |
+| Economics & business | 1,382 |
+| History & geography | 1,153 |
+| Architecture & urbanism | 1,119 |
+| Professional workshops | 1,104 |
+| Design & UX/UI | 1,087 |
+| Psychology | 1,039 |
+| Real estate marketing | 1,003 |
+| BigTech reports | 1,001 |
+| Startup pitch decks | 400 |
+
+## Construction (brief)
+
+The build pipeline producing RuDoRe is a five-stage funnel:
+
+1. **Document crawling** — topic-conditioned Yandex Search API
+   queries; PDF / PPTX / DOCX surface-document collection (2,355
+   documents / 228,902 unique pages survive deduplication and
+   mime-type verification).
+2. **Topic-aware split** — greedy multi-label document-level
+   train/test split so adjacent pages of the same deck cannot leak.
+3. **Query generation** — vision-language model conditioned on each
+   page emits a short Russian query.
+4. **Query validation** — an LLM judge labels each (query, page) pair
+   `aligned` / `not aligned` and only `aligned` pairs survive.
+5. **Per-topic quota sampling** — caps each topic at ~20k train and
+   ~2k test pairs for balance.
+6. **Rephrasing** — every accepted query is rephrased by the same VLM
+   to remove token-level overlap with the page text.
+
+The non-rephrased counterpart (`RuDoRe-NR`) is used only as a
+diagnostic control and is not published here.
+
+## Dataset precision against human judgement
+
+Two human annotators labelled a deterministic random sample of
+N = 150 judge-aligned pairs from this split.
+
+| Metric | Value |
+| --- | ---: |
+| Mean precision $\bar\pi$ | 0.89 |
+| Raw inter-annotator agreement $p_o$ | 0.94 |
+
+## License
+
+[Creative Commons Attribution 4.0](https://creativecommons.org/licenses/by/4.0/).
+The page images are screenshots taken from publicly distributed
+business documents; if you are an upstream rights-holder and want
+specific pages removed, open an issue on the dataset repo.