File size: 4,601 Bytes
a05c9c1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 | """Publish an exact, revision-pinned copy of the official single-turn IFBench data."""
import argparse
import hashlib
import json
from pathlib import Path
from urllib.request import urlopen
from datasets import Dataset, Features, Json, List, Value, load_dataset
from huggingface_hub import HfApi, hf_hub_download
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--output-dataset", required=True)
parser.add_argument("--source-revision", required=True)
args = parser.parse_args()
source = f"https://raw.githubusercontent.com/allenai/IFBench/{args.source_revision}/ifbench/data/IFBench_test.jsonl"
raw = urlopen(source).read()
records = [json.loads(line) for line in raw.splitlines() if line.strip()]
assert len(records) == 300 and len({row['key'] for row in records}) == 300
features = Features(
{
"key": Value("string"),
"prompt": Value("string"),
"instruction_id_list": List(Value("string")),
"kwargs": Json(),
}
)
dataset = Dataset.from_list(records, features=features)
assert list(dataset) == records
api = HfApi()
# Never replace an existing repository or its dataset card.
api.create_repo(args.output_dataset, repo_type="dataset", private=False, exist_ok=False)
dataset.push_to_hub(args.output_dataset, config_name="default", split="test")
generated_card = Path(hf_hub_download(args.output_dataset, "README.md", repo_type="dataset")).read_text()
yaml = generated_card.split("---", 2)[1].strip()
card = f"""---
license: odc-by
language:
- en
tags:
- evaluation
- instruction-following
{yaml}
---
| subset | description | num_samples |
| --- | --- | ---: |
| `default` | Official single-turn IFBench test prompts, unchanged; `test` split | {len(dataset)} |
# IFBench
An exact mirror of the single-turn test data from [AllenAI IFBench](https://github.com/allenai/IFBench).
Credit: Valentina Pyatkin, Saumya Malik, Victoria Graf, Hamish Ivison, Shengyi Huang, Pradeep Dasigi,
Nathan Lambert, and Hannaneh Hajishirzi, [Generalizing Verifiable Instruction Following](https://huggingface.co/papers/2507.02833).
## Provenance and license
- Source revision: `{args.source_revision}`.
- Source file: [`ifbench/data/IFBench_test.jsonl`]({source}).
- Original file SHA-256: `{hashlib.sha256(raw).hexdigest()}`.
- No filtering, prompt changes, or row reordering. All original fields and values are preserved.
- `kwargs` uses the Hugging Face Datasets JSON feature to preserve heterogeneous argument dictionaries;
use `datasets>=4.7.0` to load it as a list of dictionaries. Other fields are `key`, `prompt`, and `instruction_id_list`.
- The data is licensed under [ODC-BY-1.0](https://opendatacommons.org/licenses/by/1-0/), as stated in the
[upstream README](https://github.com/allenai/IFBench/blob/{args.source_revision}/README.md#licensing).
It is intended for research and educational use under Ai2's Responsible Use Guidelines.
It includes outputs generated by third-party models, which may be subject to separate terms.
- Verifier code is separately Apache-2.0 licensed; it is not part of this dataset release.
## Reproducing
The builder is published at `scripts/create_dataset.py`. Exact publication invocation:
```bash
python scripts/create_dataset.py --output-dataset {args.output_dataset} --source-revision {args.source_revision}
```
To reproduce independently, use an unused repository name under your own namespace. The builder refuses to overwrite
an existing repository. It requires `datasets>=4.7.0` and `huggingface_hub`, and a token with write access to the destination.
## Loading
```python
from datasets import load_dataset
dataset = load_dataset("{args.output_dataset}", "default", split="test")
```
Pin a Hub commit revision for reproducible evaluation. This repository contains evaluation data, not training data.
"""
api.upload_file(
path_or_fileobj=card.encode(), path_in_repo="README.md", repo_id=args.output_dataset, repo_type="dataset"
)
commit = api.upload_file(
path_or_fileobj=__file__,
path_in_repo="scripts/create_dataset.py",
repo_id=args.output_dataset,
repo_type="dataset",
commit_message="Publish the script that built this dataset",
)
loaded = load_dataset(args.output_dataset, "default", split="test", revision=commit.oid)
assert list(loaded) == records
print(json.dumps({"dataset": args.output_dataset, "revision": commit.oid, "samples": len(loaded)}))
if __name__ == "__main__":
main()
|