Unify README structure and add Data structure bm25 details

README.md

@@ -167,16 +167,25 @@ configs:
     path: queries/NanoCodeSearchNetRuby-*
 ---
 
-
 # NanoCodeSearchNet (with bm25 subset)
 
-A tiny, evaluation-ready slice of [CodeSearchNet](https://huggingface.co/datasets/code-search-net/code_search_net) (test set) that mirrors the spirit of [NanoBEIR](https://huggingface.co/collections/zeta-alpha-ai/nanobeir): same task, same style, but dramatically smaller so you can iterate and benchmark in minutes instead of hours.
+A lightweight, evaluation-ready subset of [CodeSearchNet](https://huggingface.co/datasets/code-search-net/code_search_net), designed for fast code-retrieval benchmarking.
+
+## What this dataset is
 
-Evaluation can be performed during and after training by integrating with Sentence Transformer's Evaluation module (InformationRetrievalEvaluator).
+- A code-retrieval benchmark with 6 programming-language splits: `Go`, `Java`, `JavaScript`, `PHP`, `Python`, `Ruby`.
+- Each split contains 50 queries and up to 10,000 code snippets.
+- Queries are function docstrings and positives are corresponding function bodies.
 
-## Performance Comparison Across Models
+## Data structure
 
-NanoCodeSearchNet Evaluation (NDCG@10)
+- `corpus`: `_id`, `text`
+- `queries`: `_id`, `text`
+- `qrels`: `query-id`, `corpus-id`, `score`
+- `bm25`: `query-id`, `corpus-ids`
+- `bm25`: A bm25 subset is included for first-stage retrieval and reranking experiments.
+
+## Performance Comparison Across Models
 
 A `bm25` subset is included for first-stage retrieval and reranking experiments.
 
@@ -186,57 +195,25 @@ A `bm25` subset is included for first-stage retrieval and reranking experiments.
 | e5-small | 0.7268 | 0.6805 | 0.7884 | 0.6386 | 0.6603 | 0.9142 | 0.6788 |
 | e5-large | 0.7836 | 0.7595 | 0.8287 | 0.6942 | 0.7291 | 0.9513 | 0.7391 |
 | bge-m3 | 0.7107 | 0.6643 | 0.6934 | 0.6138 | 0.6229 | 0.9852 | 0.6844 |
-  
-
-## What this dataset is
-
-- A collection of 6 programming-language subsets (`corpus`, `queries`, `qrels`) published on the Hugging Face Hub under `hotchpotch/NanoCodeSearchNet`.
-- Each subset contains **50 test queries** and a **corpus of up to 10,000 code snippets**.
-- Queries are function docstrings, and positives are the corresponding function bodies from the same source row.
-- Query IDs are `q-<docid>`, where `docid` is the `func_code_url` when available.
-- Built from the CodeSearchNet `test` split (`refs/convert/parquet`) with deterministic sampling (seed=42).
-- License: **Other** (see CodeSearchNet and upstream repository licenses).
 
 ## Subset names
 
-- Split names:
-  - `NanoCodeSearchNetGo`
-  - `NanoCodeSearchNetJava`
-  - `NanoCodeSearchNetJavaScript`
-  - `NanoCodeSearchNetPHP`
-  - `NanoCodeSearchNetPython`
-  - `NanoCodeSearchNetRuby`
-- Config names: `corpus`, `queries`, `qrels`
-
-## Usage
-
-```python
-from datasets import load_dataset
-
-split = "NanoCodeSearchNetPython"
-queries = load_dataset("hotchpotch/NanoCodeSearchNet", "queries", split=split)
-corpus  = load_dataset("hotchpotch/NanoCodeSearchNet", "corpus",  split=split)
-qrels   = load_dataset("hotchpotch/NanoCodeSearchNet", "qrels",   split=split)
+- Split names: `NanoCodeSearchNetGo`, `NanoCodeSearchNetJava`, `NanoCodeSearchNetJavaScript`, `NanoCodeSearchNetPHP`, `NanoCodeSearchNetPython`, `NanoCodeSearchNetRuby`
+- Config names: `corpus`, `queries`, `qrels`, `bm25`
 
-print(queries[0]["text"])
-```
+## BM25 tokenization strategy
 
-## Why Nano?
+- `transformer`: `Qwen/Qwen3-0.6B` tokenizer
+- `word_seg`: language-specific word segmentation for `ja`, `zh`, `th`, `ko`
+- `stemmer`: `PyStemmer`
+- `whitespace`: simple whitespace split (`str.split()`)
+- Auto selection chooses the best strategy per split/language using retrieval metrics.
 
-- **Fast eval loops**: 50 queries × 10k docs fits comfortably on a single GPU/CPU run.
-- **Reproducible**: deterministic sampling and stable IDs.
-- **Drop-in**: BEIR/NanoBEIR-style schemas, so existing IR loaders need minimal tweaks.
+## Upstream source
 
-### Upstream sources
-
-- Original data: **CodeSearchNet** — [CodeSearchNet Challenge: Evaluating the State of Semantic Code Search: 1909.09436](https://huggingface.co/papers/1909.09436).
-- Base dataset: [code-search-net/code_search_net](https://huggingface.co/datasets/code-search-net/code_search_net) (Hugging Face Hub).
-- Inspiration: **NanoBEIR** (lightweight evaluation subsets).
+- Original dataset: [code-search-net/code_search_net](https://huggingface.co/datasets/code-search-net/code_search_net)
+- Paper: [CodeSearchNet Challenge (1909.09436)](https://huggingface.co/papers/1909.09436)
 
 ## License
 
-Other. This dataset is derived from CodeSearchNet and ultimately from open-source GitHub repositories. Please respect original repository licenses and attribution requirements.
-
-## Author
-
-- Yuichi Tateno
+Other. This dataset is derived from CodeSearchNet and follows upstream repository licensing and attribution requirements.