Add leaderboard database README

Add an English dataset card explaining that the leaderboard database is a DuckDB transformation of the raw hakari-bench/results artifacts, how it is built from the GitHub scripts, how the leaderboard consumes it, and where to find the official result submission workflow.

README.md

@@ -0,0 +1,90 @@
+---
+pretty_name: HAKARI-Bench Leaderboard Database
+tags:
+- information-retrieval
+- retrieval
+- benchmark
+- duckdb
+- leaderboard
+- evaluation
+---
+
+# HAKARI-Bench Leaderboard Database
+
+This dataset publishes the DuckDB database used by the HAKARI-Bench
+leaderboard. It is a transformed, leaderboard-ready view of the raw benchmark
+result files stored in [`hakari-bench/results`](https://huggingface.co/datasets/hakari-bench/results).
+
+The raw results dataset is the source of truth: it stores per-task compressed
+JSON result artifacts generated by HAKARI-Bench. This dataset repackages those
+raw artifacts into `duckdb/hakari_bench.duckdb` so the leaderboard, notebooks,
+and SQL-based analysis can query model scores, task metadata, variants,
+runtime metadata, and precomputed viewer tables efficiently.
+
+## Contents
+
+- `duckdb/hakari_bench.duckdb`: the current HAKARI-Bench leaderboard DuckDB
+  warehouse.
+
+The database schema and leaderboard query conventions are documented in the
+GitHub repository:
+
+- [DuckDB schema and leaderboard query guide](https://github.com/hakari-bench/hakari-bench/blob/main/docs/duckdb_schema.md)
+- [Leaderboard metric policy](https://github.com/hakari-bench/hakari-bench/blob/main/docs/leaderboard_metrics.md)
+
+## How This Dataset Is Built
+
+The DuckDB file is generated by scripts in
+[`hakari-bench/hakari-bench`](https://github.com/hakari-bench/hakari-bench).
+The usual build path syncs the raw results dataset and rebuilds the database:
+
+```bash
+uv run python scripts/sync_remote_results_and_rebuild.py --sync-backend xet
+```
+
+The lower-level builder can also create or append a database from local result
+directories:
+
+```bash
+uv run python scripts/build_results_database_and_report.py \
+  --results-dir output/hakari-results \
+  --duckdb-path output/hakari-results/hakari_bench.duckdb
+```
+
+See the [evaluation runbook](https://github.com/hakari-bench/hakari-bench/blob/main/docs/evaluation_runbook.md)
+for the full rebuild, append, cache, and viewer workflow.
+
+## How It Is Used
+
+The public leaderboard consumes this dataset:
+
+- [HAKARI-Bench Leaderboard Space](https://huggingface.co/spaces/hakari-bench/leaderboard)
+
+The HAKARI-Bench viewer can download the DuckDB directly from this dataset:
+
+```bash
+uv run hakari-bench web \
+  --hf-dataset-repo-id hakari-bench/leaderboard_database \
+  --hf-dataset-path duckdb/hakari_bench.duckdb
+```
+
+## Contributing Official Results
+
+To add model evaluation results to the official HAKARI-Bench data flow, submit
+the raw `.json.xz` result files to
+[`hakari-bench/results`](https://huggingface.co/datasets/hakari-bench/results).
+After raw results are reviewed and merged, this leaderboard database can be
+rebuilt from those files.
+
+Use the GitHub workflow for evaluating a new model and preparing an official
+results submission:
+
+- [New model evaluation to results PR workflow](https://github.com/hakari-bench/hakari-bench/blob/main/docs/new_model_results_workflow.md)
+- [Contributing result files](https://github.com/hakari-bench/hakari-bench/blob/main/docs/contributing_results.md)
+
+## Related Resources
+
+- [HAKARI-Bench GitHub repository](https://github.com/hakari-bench/hakari-bench)
+- [Raw result artifacts dataset](https://huggingface.co/datasets/hakari-bench/results)
+- [Leaderboard database dataset](https://huggingface.co/datasets/hakari-bench/leaderboard_database)
+- [Leaderboard Space](https://huggingface.co/spaces/hakari-bench/leaderboard)