Update dataset card to reflect filtered release and noise-filter rules

README.md

@@ -210,73 +210,48 @@ license: apache-2.0
 
 # cornstack-samples
 
-🚧 This dataset is under active development and may change.
-
-
-
-`cornstack-samples` is a compact, sampled version of CoRNStack for code-focused embedding training, reranker training, and retrieval experiments.
+Filtered CoRNStack sample subsets for code retrieval training.
 
 Source dataset and paper:
 - CoRNStack collection: https://huggingface.co/collections/nomic-ai/cornstack
 - CoRNStack paper: https://huggingface.co/papers/2412.01007
 
-## Why this dataset
+## What This Release Contains
 
-CoRNStack is large. This dataset provides smaller, practical subsets for faster iteration and lower-cost experiments.
+This release keeps the original subset layout (6 languages x pair + hard-negatives) and applies a deterministic rule-based noise filter.
 
-Important notes:
-- This is a sampled dataset, not the full CoRNStack distribution.
-- If you need maximum coverage or full-scale training, use the original CoRNStack collection.
+Important:
+- Counts are post-filter counts, so they are slightly smaller than the nominal 2M / 100k targets.
+- Data is published in normalized IR format.
 
-## Config layout
+## Config Layout And Schema
 
-Each language is published as **two separate configs**:
+Each language is published as two configs with split `train`:
 - `{lang}-v1-pair-2M`
 - `{lang}-v1-hard-negatives-100k`
 
-Each config uses split `train`.
-
-### `{lang}-v1-pair-2M`
-Fields:
-- `query`
-- `document`
-
-Target rows per language: ~2,000,000
+Schema:
+- Pair configs: `query`, `pos`
+- Hard-negative configs: `query`, `pos`, `negs` (list[string])
 
-### `{lang}-v1-hard-negatives-100k`
-Fields:
-- `query`
-- `positive`
-- `negative_1`
-- `negative_2`
-- `negative_3`
-- `negative_4`
-- `negative_5`
-- `negative_6`
-- `negative_7`
+## Subsets And Row Counts (Post-filter)
 
-Target rows per language: ~100,000
+| Subset (config name) | split | num_examples |
+| --- | --- | ---: |
+| `go-v1-pair-2M` | `train` | 1,992,985 |
+| `go-v1-hard-negatives-100k` | `train` | 99,663 |
+| `java-v1-pair-2M` | `train` | 1,752,593 |
+| `java-v1-hard-negatives-100k` | `train` | 87,504 |
+| `javascript-v1-pair-2M` | `train` | 1,960,276 |
+| `javascript-v1-hard-negatives-100k` | `train` | 98,025 |
+| `php-v1-pair-2M` | `train` | 1,710,537 |
+| `php-v1-hard-negatives-100k` | `train` | 85,460 |
+| `python-v1-pair-2M` | `train` | 1,990,051 |
+| `python-v1-hard-negatives-100k` | `train` | 99,535 |
+| `ruby-v1-pair-2M` | `train` | 1,583,047 |
+| `ruby-v1-hard-negatives-100k` | `train` | 79,040 |
 
-## Complete subset list (exact)
-
-All subsets are published with split `train`.
-
-| Subset (config name) | split | num_examples | type |
-| --- | --- | ---: | --- |
-| `go-v1-pair-2M` | `train` | 2,000,000 | pair |
-| `go-v1-hard-negatives-100k` | `train` | 100,000 | hard-negatives |
-| `java-v1-pair-2M` | `train` | 2,000,000 | pair |
-| `java-v1-hard-negatives-100k` | `train` | 100,000 | hard-negatives |
-| `javascript-v1-pair-2M` | `train` | 2,000,000 | pair |
-| `javascript-v1-hard-negatives-100k` | `train` | 100,000 | hard-negatives |
-| `php-v1-pair-2M` | `train` | 2,000,000 | pair |
-| `php-v1-hard-negatives-100k` | `train` | 100,000 | hard-negatives |
-| `python-v1-pair-2M` | `train` | 2,000,000 | pair |
-| `python-v1-hard-negatives-100k` | `train` | 100,000 | hard-negatives |
-| `ruby-v1-pair-2M` | `train` | 2,000,000 | pair |
-| `ruby-v1-hard-negatives-100k` | `train` | 100,000 | hard-negatives |
-
-## Quick usage
+## Quick Usage
 
 ```python
 from datasets import load_dataset
@@ -288,26 +263,48 @@ print(pair_ds.column_names, len(pair_ds))
 print(hard_ds.column_names, len(hard_ds))
 ```
 
-## Creation process (rough)
-
-For each language:
-1. Read CoRNStack shard files (`shard-*.jsonl.gz`).
-2. Count lines per shard and allocate per-shard targets.
-3. Build pair data via random line-index sampling with valid string `query`/`document`.
-4. Build hard-negative data from disjoint rows:
-   - valid string `query`/`document`
-   - `negatives` contains at least 7 unique non-empty strings
-   - sample 7 negatives per row
-   - cap output via reservoir sampling
-
-Default seed is fixed (`42`) for reproducibility under identical source snapshots and parameters.
-
 ## License
 
 This dataset follows CoRNStack and is released under **Apache-2.0**.
 
-## Citation and attribution
+## Citation And Attribution
 
 If you use this dataset, please cite and attribute CoRNStack:
 - Paper: https://huggingface.co/papers/2412.01007
 - Collection: https://huggingface.co/collections/nomic-ai/cornstack
+
+## Noise Filtering Algorithm (Rule-based)
+
+The following deterministic rules are applied before publishing this release.
+
+1. Prefix-based noisy query removal
+A row is dropped if `query` starts with any of the following prefixes:
+- `TODO`
+- `GET /`
+- `POST /`
+- `PUT /`
+- `DELETE /`
+- `Display a listing of the resource.`
+- `Store a newly created resource in storage.`
+- `Show the form for editing the specified resource.`
+- `Update the specified resource in storage.`
+- `Show the form for creating a new resource.`
+- `Remove the specified resource from storage.`
+- `Display the specified resource.`
+- `Transform the resource into an array.`
+- `Autogenerated method stub`
+- `Auto generated`
+- `this down() migration is autogenerated`
+- `this up() migration is autogenerated`
+- `"/ renamed from:"`
+- `"/ access modifiers changed from:"`
+
+2. Minimum positive-document length
+A row is dropped if the positive side is shorter than 30 characters.
+- Pair task: `document` length >= 30 required
+- Hard-negatives task: `positive` length >= 30 required
+
+3. Hard-negative validity constraint
+For hard-negative configs, at least one valid negative must remain after normalization (`min_negs = 1`).
+
+This filtering is purely rule-based (no model scoring), targeting high-noise templates and low-information positives while preserving broad retrieval coverage.