---
license: mit
language:
- en
size_categories:
- 1M<n<10M
task_categories:
- text-classification
configs:
- config_name: default
  data_files:
  - split: compliance_clean
    path: data/compliance_clean/*.parquet
  - split: compliance
    path:
    - data/compliance_clean/*.parquet
    - data/compliance_flagged/*.parquet
  - split: noncompliance
    path: data/noncompliance/*.parquet
  - split: ambiguous
    path: data/ambiguous/*.parquet
- config_name: labels
  data_files:
  - split: train
    path: data/labels/*.parquet
---

# CompOrca

> **Paper:** [CompOrca: Corpus-Scale Compliance Labelling of Instruction-Tuning Data](https://arxiv.org/abs/2609.37807)  
> Philipp E. Glass and Alina Miron. PlurVA-LLM Workshop @ AACL-IJCNLP 2026.

> **Purpose:** A training-scale corpus of instruction-response examples for
> studying model compliance and noncompliance.

Every row of OpenOrca, labelled COMPLIANCE or NONCOMPLIANCE. Each row was
classified five times independently by an LLM judge. Rows where the five
passes did not agree are kept as `ambiguous`.

An additional `compliance_clean` split contains unanimous-compliance rows that
WildGuard also classified as compliant.

Source: [`Open-Orca/OpenOrca`](https://huggingface.co/datasets/Open-Orca/OpenOrca),
split `train` at revision [`e9c87b4`](https://huggingface.co/datasets/Open-Orca/OpenOrca/tree/e9c87b4abb2609913751f9b26553fdb9c061796c) (4,233,923 rows).

| split | rows | |
|---|---:|---|
| `compliance_clean` | 3,769,249 | all five passes and WildGuard COMPLIANCE |
| `noncompliance` | 54,004 | all five passes NONCOMPLIANCE |
| `ambiguous` | 168,157 | the passes disagreed |
| `compliance` | 4,011,762 | all five passes COMPLIANCE (composed of `compliance_clean` + `compliance_flagged`) |

Columns: `idx`, `id`, `system_prompt`, `question`, `response`, `label` —
OpenOrca's columns plus `idx`, the row's position in the upstream split, and
`label`, constant within each split (`COMPLIANCE`, `NONCOMPLIANCE`, and null
respectively).

The recommended splits are `compliance_clean` and `noncompliance`. The
`compliance` split is provided as a composite split combining `compliance_clean`
and `compliance_flagged` (242,513 rows) without duplicating text data on disk;
`ambiguous` is included for completeness.

```python
from datasets import load_dataset
comp = load_dataset("cemiu/CompOrca", split="compliance_clean")
noncomp = load_dataset("cemiu/CompOrca", split="noncompliance")
```

The `labels` config carries the labels for all 4,233,923 rows without the
text: `idx`, `label` (`COMPLIANCE`, `NONCOMPLIANCE`, or null when the passes
disagreed), `votes` (number of NONCOMPLIANCE votes, 0–5), and `pass_1` …
`pass_5` (the individual judgments), and `wildguard_label` (WildGuard's
response-refusal output: `COMPLIANCE`, `REFUSAL`, or `N/A`).

Human-annotated evaluation rows: [`cemiu/CompOrca-gold`](https://huggingface.co/datasets/cemiu/CompOrca-gold).

MIT, as with the underlying OpenOrca text.

## Citation

```bibtex
@misc{glass2026comporca,
  title         = {{CompOrca}: Corpus-Scale Compliance Labelling of Instruction-Tuning Data},
  author        = {Glass, Philipp E. and Miron, Alina},
  year          = {2026},
  eprint        = {2609.37807},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.37807}
}
```