---
language:
  - en
license: cc-by-4.0
task_categories:
  - text-classification
pretty_name: Dedup-Filtered HelpSteer3 with Rubrics
size_categories:
  - 10K<n<100K
source_datasets:
  - nvidia/HelpSteer3
tags:
  - reward-model
  - preference
  - rubric
  - sdpo
---

# Dedup-Filtered HelpSteer3 (+ rubrics on train)

A cleaned version of [`nvidia/HelpSteer3`](https://huggingface.co/datasets/nvidia/HelpSteer3)
(`preference` config) with three modifications:

1. **Filter** out `domain == "multilingual"` and `overall_preference == 0` (tie) rows.
2. **Within-split dedup** by content hash (`sha1(context, response1, response2)`),
   keeping the first occurrence. Upstream HS3 has ~35% byte-identical duplicate
   rows after the filter step.
3. **Cross-split dedup**: drop validation rows whose content hash also appears
   in train. Upstream HS3 has ~918 of ~1,553 filtered-deduped val rows that are
   byte-identical to a train row, which would otherwise cause train/val leakage.

Both splits carry a `rubric` column. **Train rubrics were generated by
`claude-opus-4-6-v1`** via an external multicall rubric-gen pipeline and then
content-hash joined onto each train row. Each rubric is a judging rubric for
the pair (sections: `Task`, `Hard requirements`, `Discriminative criteria`,
`Pitfalls`). Train rubrics are guaranteed non-empty (100% coverage).
**Validation** rubrics are always the empty string — val eval does not use
the rubric; the column is present only for HF DatasetDict schema parity.

## Sizes

| Split | Rows | `rubric` |
|---|---|---|
| train | 19557 | non-empty (100% covered) |
| validation | 635 | empty string |

## Schema

`train` and `validation` both inherit the upstream HS3 `preference` schema
(`context`, `response1`, `response2`, `overall_preference`,
`individual_preference`, `domain`, …). `train` adds one column:

- `rubric: str` — evaluation rubric for the pair, produced by the rubric-gen
  pipeline (sections: `Task`, `Hard requirements`, `Discriminative criteria`,
  `Pitfalls`). Guaranteed non-empty.

## Dedup details

- Content hash: `sha1(json.dumps([context, response1, response2], sort_keys=True, ensure_ascii=False))`.
- Keep-first on first occurrence. Upstream HS3 has 13 groups where
  `(context, response1, response2)` are identical but `overall_preference` /
  `individual_preference` differ; for those groups, this dataset keeps the
  first occurrence's labels.

## Train/val leak

The raw HS3 `preference` config has a substantial train/val content overlap:
after standard filters, 918 of 1,553 unique val pairs are also in train.
This dataset removes those from val, leaving 635 genuinely-unseen pairs.

## License

Inherits upstream HelpSteer3's license. Please cite the original dataset:

```
@misc{wang2025helpsteer3,
  title={HelpSteer 3},
  author={Wang, Zhilin and others},
  year={2025},
  publisher={NVIDIA},
  url={https://huggingface.co/datasets/nvidia/HelpSteer3}
}
```