README.md

---
license: cc-by-nc-4.0
language:
- en
task_categories:
- text-generation
tags:
- code
- non-commercial
- sft
- chatml
pretty_name: CodeAlpaca OpenAI Native
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: test
    path: data/test.parquet
---

# CodeAlpaca OpenAI Native

This is a deterministic, lossless-formatting derivative of
[`sahil2801/CodeAlpaca-20k`](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k),
modernized with a typed OpenAI/TRL `messages` column and decontaminated against
the HumanEval and MBPP test sets. The original Alpaca columns remain available
for backward compatibility.

## Intended use

```python
from datasets import load_dataset
from trl import SFTTrainer

dataset = load_dataset("Archangel-system/codealpaca-openai-native")
trainer = SFTTrainer(model="your-model", train_dataset=dataset["train"])
```

No row-level JSON parsing or prompt concatenation is required. Choose a model
whose tokenizer has an appropriate chat template.

## Schema and conversion rule

Each row has:

- `messages`: a typed list of two records, strictly `user` then `assistant`;
- `instruction`, `input`, and `output`: the unchanged source strings.

The user message equals `instruction` when `input.strip()` is empty. Otherwise
it equals `instruction + "\n\n" + input` (exactly two newline characters).
The assistant message equals `output`. No answer was rewritten or generated.

## Source audit

The build downloaded the source's Hub auto-converted Parquet shard directly.

| Measurement | Count |
|---|---:|
| Source rows | 20022 |
| Non-empty `input` rows | 10258 |
| Empty `output` rows | 6 |
| Exact duplicate instruction rows | 0 |
| Aggressively normalized duplicate instruction rows | 0 |
| Exact `output == instruction` rows | 0 |
| Normalized `output == instruction` rows | 0 |

## Decontamination

Two criteria are applied to CodeAlpaca `instruction` strings against benchmark
**test** prompts. The benchmark sets are never modified.

1. Exact normalized equality: `re.sub(r"[^a-z0-9]+", "", s.lower())`.
2. Shared 13-gram after lowercasing and tokenizing with `[a-z0-9]+`.

| Benchmark | Test items | Exact-normalized source rows | 13-gram source rows |
|---|---:|---:|---:|
| HumanEval | 164 | 0 | 0 |
| MBPP (`full`, `test`, field `text`) | 500 | 1 | 2 |

Any source row matching either criterion was removed before splitting. Complete
row-to-benchmark evidence is in `rejections.jsonl`. A row touching multiple
items or criteria is rejected once but retains every hit in that log.

## Filtering, deduplication, and split

| Stage / reason | Rows removed |
|---|---:|
| Empty output | 6 |
| Benchmark contamination (union) | 2 |
| Duplicate normalized `(instruction, input)` | 0 |
| Train/test signature purge from train | 0 |
| **Total removed** | **8** |
| **Rejection rate** | **0.0400%** |

Rows are deduplicated on the tuple `(normalize_text(instruction),
normalize_text(input))`, where `normalize_text` is the same exact-normalization
function used by verification. The first source occurrence is retained.

The remaining rows are shuffled with Python's `random.Random(42)`. The test
size is `min(1000, round(0.05 * n))`; test receives the first shuffled indices.
Any normalized pair crossing train/test would be purged from **train**, never
test. The measured post-build intersection is 0.

| Split | Rows |
|---|---:|
| train | 19014 |
| test | 1000 |
| total | 20014 |

## Lengths and diversity

Lengths below are lexical token counts using `[a-z0-9]+`.

| Field | Min | Median | Mean | P95 | Max |
|---|---:|---:|---:|---:|---:|
| User message | 4 | 16.0 | 17.5716 | 31.0 | 110 |
| Assistant message | 0 | 20.0 | 28.4568 | 82.0 | 633 |

The final dataset contains 20014
exactly unique instructions and 20014
unique aggressively normalized `(instruction, input)` pairs. It preserves
10252 rows with non-empty inputs.
No semantic-quality filter, language classifier, code execution, or model-based
judgment was applied; this release intentionally does not rewrite source data.

## Reproducibility and files

- Seed: `42`.
- Build script: `build.py`.
- Independent Hub verification script: `verify.py`.
- Machine-readable measurements: `stats.json`.
- Rejection evidence: `rejections.jsonl`.
- Validation report: `validation_report.json`.
- Source revision: `sahil2801/CodeAlpaca-20k@152bb5e9a29651266b018106053980070a0521a1`.
- HumanEval revision: `openai/openai_humaneval@7dce6050a7d6d172f3cc5c32aa97f52fa1a2e544`.
- MBPP revision: `google-research-datasets/mbpp@4bb6404fdc6cacfda99d4ac4205087b89d32030c`.

| File | SHA-256 |
|---|---|
| Source Parquet as downloaded | `5a2fa0c510d42843292c583f357086abb59f95796e0f55727b97e449aa93bed9` |
| `data/train.parquet` | `c00b9902c25983f60ebadb0a08905f42d44c5954ca0c8d5b4e0044e527adac5e` |
| `data/test.parquet` | `1200ef59fdc5854a2c571857c3695488754a783b6e17d778e392f05f60455864` |
| `rejections.jsonl` | `bf88205147fb3074152c43849408aa9476f44441d8f91b99d213980b7d0c3ac7` |

Run `python build.py` to rebuild locally or `python build.py --publish` to rebuild
and upload. Publication reads `HF_TOKEN` from the environment or from the
profile `.env`; the token is never written to an artifact.

## License and attribution

**This dataset is CC BY-NC 4.0 — non-commercial use only.**

The upstream repository [`sahil280114/codealpaca`](https://github.com/sahil280114/codealpaca)
ships **two separate licence files**, and they do not say the same thing:

| File | Covers | Licence |
|---|---|---|
| `LICENSE` | the training/generation **code** | Apache-2.0 |
| `DATA_LICENSE` | `data/code_alpaca_20k.json`, i.e. **the data used here** | **CC BY-NC 4.0** |

The Hugging Face mirror [`sahil2801/CodeAlpaca-20k`](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k)
is tagged `cc-by-4.0`, which **drops the NonCommercial clause** present in the upstream
`DATA_LICENSE`. We verified both files directly rather than trusting the mirror's tag, and
this derivative follows the stricter upstream data licence: **CC BY-NC 4.0**.

Practical consequence: you may use, modify and redistribute this dataset **for
non-commercial purposes with attribution**. Training a model on it for a commercial
product is **not** covered. If you need a commercially usable code-instruction set,
this is not the right dataset.

Attribution:
- Data: Sahil Chaudhary, *Code Alpaca*, `sahil280114/codealpaca`, `DATA_LICENSE` (CC BY-NC 4.0).
- Outputs were generated with OpenAI `text-davinci-003`; OpenAI's terms of use apply to
  the generated content independently of this licence.

## Limitations

The original outputs were generated with `text-davinci-003`; this release does
not independently validate their correctness, security, language, or
executability. Decontamination is limited to exact normalized equality and
lexical 13-gram overlap against the pinned HumanEval and MBPP test prompts; it
does not prove absence of semantic paraphrases. The test split measures this
source distribution and is not a replacement for external code benchmarks.