README.md

---
dataset_info:
- config_name: go-v1-hard-negatives-100k
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  - name: negs
    list: string
  splits:
  - name: train
    num_examples: 87647
- config_name: go-v1-pair-2M
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  splits:
  - name: train
    num_examples: 1541111
- config_name: java-v1-hard-negatives-100k
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  - name: negs
    list: string
  splits:
  - name: train
    num_examples: 81657
- config_name: java-v1-pair-2M
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  splits:
  - name: train
    num_examples: 1491655
- config_name: javascript-v1-hard-negatives-100k
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  - name: negs
    list: string
  splits:
  - name: train
    num_examples: 79684
- config_name: javascript-v1-pair-2M
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  splits:
  - name: train
    num_examples: 1310965
- config_name: php-v1-hard-negatives-100k
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  - name: negs
    list: string
  splits:
  - name: train
    num_examples: 75632
- config_name: php-v1-pair-2M
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  splits:
  - name: train
    num_examples: 1343442
- config_name: python-v1-hard-negatives-100k
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  - name: negs
    list: string
  splits:
  - name: train
    num_examples: 97147
- config_name: python-v1-pair-2M
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  splits:
  - name: train
    num_examples: 1807480
- config_name: ruby-v1-hard-negatives-100k
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  - name: negs
    list: string
  splits:
  - name: train
    num_examples: 68382
- config_name: ruby-v1-pair-2M
  features:
  - name: query
    dtype: string
  - name: pos
    dtype: string
  splits:
  - name: train
    num_examples: 1175219
configs:
- config_name: go-v1-hard-negatives-100k
  data_files:
  - split: train
    path: go-v1-hard-negatives-100k/train-*
- config_name: go-v1-pair-2M
  data_files:
  - split: train
    path: go-v1-pair-2M/train-*
- config_name: java-v1-hard-negatives-100k
  data_files:
  - split: train
    path: java-v1-hard-negatives-100k/train-*
- config_name: java-v1-pair-2M
  data_files:
  - split: train
    path: java-v1-pair-2M/train-*
- config_name: javascript-v1-hard-negatives-100k
  data_files:
  - split: train
    path: javascript-v1-hard-negatives-100k/train-*
- config_name: javascript-v1-pair-2M
  data_files:
  - split: train
    path: javascript-v1-pair-2M/train-*
- config_name: php-v1-hard-negatives-100k
  data_files:
  - split: train
    path: php-v1-hard-negatives-100k/train-*
- config_name: php-v1-pair-2M
  data_files:
  - split: train
    path: php-v1-pair-2M/train-*
- config_name: python-v1-hard-negatives-100k
  data_files:
  - split: train
    path: python-v1-hard-negatives-100k/train-*
- config_name: python-v1-pair-2M
  data_files:
  - split: train
    path: python-v1-pair-2M/train-*
- config_name: ruby-v1-hard-negatives-100k
  data_files:
  - split: train
    path: ruby-v1-hard-negatives-100k/train-*
- config_name: ruby-v1-pair-2M
  data_files:
  - split: train
    path: ruby-v1-pair-2M/train-*
license: apache-2.0
---
# cornstack-samples

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

## What This Release Contains

This release keeps the original subset layout (6 languages x pair + hard-negatives) and applies deterministic rule-based filtering.

In this revision, query-level deduplication is also applied per subset: if `query` is duplicated, only the first row is kept.

## Config Layout And Schema

Each language is published as two configs with split `train`:
- `{lang}-v1-pair-2M`
- `{lang}-v1-hard-negatives-100k`

Schema:
- Pair configs: `query`, `pos`
- Hard-negative configs: `query`, `pos`, `negs` (list[string])

## Subsets And Row Counts (After Filter + Query Dedup)

| Subset (config name) | split | num_examples |
| --- | --- | ---: |
| `go-v1-pair-2M` | `train` | 1,541,111 |
| `go-v1-hard-negatives-100k` | `train` | 87,647 |
| `java-v1-pair-2M` | `train` | 1,491,655 |
| `java-v1-hard-negatives-100k` | `train` | 81,657 |
| `javascript-v1-pair-2M` | `train` | 1,310,965 |
| `javascript-v1-hard-negatives-100k` | `train` | 79,684 |
| `php-v1-pair-2M` | `train` | 1,343,442 |
| `php-v1-hard-negatives-100k` | `train` | 75,632 |
| `python-v1-pair-2M` | `train` | 1,807,480 |
| `python-v1-hard-negatives-100k` | `train` | 97,147 |
| `ruby-v1-pair-2M` | `train` | 1,175,219 |
| `ruby-v1-hard-negatives-100k` | `train` | 68,382 |

Total rows:
- Pair: 8,669,872
- Hard-negatives: 490,149
- Overall: 9,160,021

## Filter Impact (Query Dedup Stage)

The table below shows only the query-dedup impact on top of the previous rule-based filter.

| Subset | before | after | removed | removed_ratio |
| --- | ---: | ---: | ---: | ---: |
| `go-v1-pair-2M` | 1,992,985 | 1,541,111 | 451,874 | 22.67% |
| `go-v1-hard-negatives-100k` | 99,663 | 87,647 | 12,016 | 12.06% |
| `java-v1-pair-2M` | 1,752,593 | 1,491,655 | 260,938 | 14.89% |
| `java-v1-hard-negatives-100k` | 87,504 | 81,657 | 5,847 | 6.68% |
| `javascript-v1-pair-2M` | 1,960,276 | 1,310,965 | 649,311 | 33.12% |
| `javascript-v1-hard-negatives-100k` | 98,025 | 79,684 | 18,341 | 18.71% |
| `php-v1-pair-2M` | 1,710,537 | 1,343,442 | 367,095 | 21.46% |
| `php-v1-hard-negatives-100k` | 85,460 | 75,632 | 9,828 | 11.50% |
| `python-v1-pair-2M` | 1,990,051 | 1,807,480 | 182,571 | 9.17% |
| `python-v1-hard-negatives-100k` | 99,535 | 97,147 | 2,388 | 2.40% |
| `ruby-v1-pair-2M` | 1,583,047 | 1,175,219 | 407,828 | 25.76% |
| `ruby-v1-hard-negatives-100k` | 79,040 | 68,382 | 10,658 | 13.48% |

## Quick Usage

```python
from datasets import load_dataset

pair_ds = load_dataset("hotchpotch/cornstack-samples", "python-v1-pair-2M", split="train")
hard_ds = load_dataset("hotchpotch/cornstack-samples", "python-v1-hard-negatives-100k", split="train")

print(pair_ds.column_names, len(pair_ds))
print(hard_ds.column_names, len(hard_ds))
```

## License

This dataset follows CoRNStack and is released under **Apache-2.0**.

## 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 text is shorter than 30 characters.
- Pair configs: `pos` length >= 30 required
- Hard-negative configs: `pos` length >= 30 required

3. Hard-negative validity constraint
For hard-negative configs, at least one valid negative must remain after normalization (`min_negs = 1`).

4. Query-level deduplication
Within each subset split, rows are grouped by exact `query` string.
- Keep the first occurrence
- Drop all later duplicates

This filtering is purely rule-based (no model scoring), targeting high-noise templates and low-information positives while preserving broad retrieval coverage.