Blame for README.md

---
license: apache-2.0
language:
- en
size_categories:
- 10M<n<100M
task_categories:
- text-generation
tags:
- preference
- reward-model
- rlhf
- criteria
- nectar
pretty_name: C3PO criterion-conditioned preference pairs
configs:
- config_name: criterion
  data_files:
  - split: train
    path: criterion/train.parquet
  - split: validation
    path: criterion/valid.parquet
  - split: test
    path: criterion/test.parquet
- config_name: holistic
  data_files:
  - split: train
    path: holistic/train.parquet
  - split: validation
    path: holistic/valid.parquet
  - split: test
    path: holistic/test.parquet
---

# C3PO: criterion-conditioned preference pairs from Nectar

Anonymous release accompanying a paper under review.

This dataset relabels the pairwise preferences of
[Nectar](https://huggingface.co/datasets/berkeley-nest/Nectar) under a fixed set of
fifteen named criteria. No new text is generated: every prompt and every response is
taken verbatim from Nectar. What is added is, for each pair, the criterion under which
it was judged and the preference label obtained under that criterion.

Two configurations are released, built from the same prompts by the same judge:

| Config | Pairs | Train | Validation | Test | What it is |
|---|---|---|---|---|---|
| `criterion` | 8,433,953 | 7,592,311 | 420,748 | 420,894 | every pair with a consensus winner under an assigned criterion |
| `holistic` | 2,912,569 | 2,621,420 | 145,657 | 145,492 | the same judge asked for the better response overall, with no criterion |

The `holistic` config is the control: same prompts, same candidate pool, same judge,
same two-order protocol, differing only in the absence of a criterion.

## How the pairs were made

1. **Criterion assignment.** For each Nectar prompt, the judge picks the three criteria
   from the fixed set of fifteen (below) that best suit the prompt.
2. **Pairwise judging.** For each assigned criterion, every unordered pair among the
   prompt's seven candidate responses is shown to the judge, which picks the response
   that better satisfies that criterion. Each pair is judged twice, once in each
   presentation order.
3. **Consensus.** A pair is kept only if the judge names the same winner in both orders.
   Pairs where the two orders disagree, or where the judge declines in either order,
   are dropped.
4. **Holistic pass.** The same procedure is run once more per pair with no criterion,
   asking for the better response overall. This produces the `holistic` config.
5. **Cleaning.** Nectar sometimes lists the same response text in two of a prompt's seven
   slots. Pairs are formed over slots, so such prompts produce text-identical rows, and
   occasionally a pair whose two sides are the same text. The release drops rows whose
   `chosen` equals `rejected` and exact duplicates on (`prompt`, `chosen`, `rejected`,
   `criterion`), keeping the first occurrence: 2.4% of the criterion rows and 2.6% of the
   holistic rows. Every row in the release is therefore unique and has two distinct responses.

The judge for every call is `Qwen3-30B-A3B-Instruct-2507`, decoded greedily.

### The fifteen criteria

| Criterion | Definition shown to the judge |
|---|---|
| clarity | The response is clear, well-written, and easy to understand. |
| correctness | The response is factually accurate and free of errors. |
| relevance | The response directly addresses the user's question or request. |
| completeness | The response thoroughly covers all aspects of the query. |
| helpfulness | The response provides actionable, useful information. |
| safety | The response avoids harmful, dangerous, or unethical content. |
| reasoning | The response demonstrates sound logical reasoning. |
| creativity | The response shows originality and creative thinking where appropriate. |
| conciseness | The response is appropriately brief without sacrificing important details. |
| tone | The response uses an appropriate, respectful, and professional tone. |
| nuanced explanation | Provides detailed, balanced, and context-specific explanations that account for multiple perspectives and complexities. |
| technical specificity | Evaluates how precisely and accurately a response addresses the technical aspects of a given topic or problem. |
| practical applicability | Evaluates how well a response provides actionable and feasible advice that can be applied in real-world situations. |
| consequence awareness | The extent to which a response acknowledges or ignores potential consequences, risks, or downsides of a particular action or decision. |
| contextual understanding | Evaluates how well a response demonstrates comprehension of the underlying context, motivations, and implications of a given statement or situation. |

## Schema

Both configs share one schema.

| Column | Type | Meaning |
|---|---|---|
| `prompt` | string | the Nectar prompt, unchanged |
| `chosen` | string | the response the judge preferred, verbatim from Nectar |
| `rejected` | string | the other response, verbatim from Nectar |
| `criterion` | string | the criterion the pair was judged under; the empty string in the `holistic` config |
| `flipped` | bool | `true` when the label disagrees with Nectar's original GPT-4 ranking of the two responses |

The criterion is kept as a column rather than written into the prompt, so that users
can choose their own conditioning template. In the accompanying paper the criterion
enters the prompt as a system turn, `You are an assistant that values {criterion}.`,
and the holistic control uses `you are a helpful assistant`.

## Splits

The split is by prompt: a prompt and all of its pairs fall into exactly one of train,
validation and test, and the two configs share the same prompt-level split (90/5/5,
seed 42). A model trained on `criterion/train` has therefore never seen a
`holistic/test` prompt, and vice versa.

## Loading

```python
from datasets import load_dataset

crit = load_dataset("C3PO-Project/c3po", "criterion")
hol  = load_dataset("C3PO-Project/c3po", "holistic")

row = crit["train"][0]
system = f"You are an assistant that values {row['criterion']}."
```

## Source and license

Prompts and responses come from Nectar (Zhu et al., 2024), released under Apache 2.0,
and this dataset is released under the same license. The labels and the criterion
assignments are new.

## Citation

A citation will be added when the accompanying paper is public.