---
license: mit
task_categories:
- text-classification
- text-ranking
language:
- en
pretty_name: OCN Reward Preference Pairs
---

# OCN Reward Preference Pairs

This dataset contains counterfactual response pairs derived from lexical OCN candidates in the main Gemma 4 E2B and Qwen 3.5 2B generation experiment. Each detected source response is paired with a direct affirmative rewrite generated by `Qwen/Qwen3.5-2B`.

The source dataset is `ritwikraha/ocn-empty-negations-detection-main-gemma4-qwen35`.

## Variant Types

- `candidate_ocn`: the original model response containing at least one lexical OCN candidate.
- `plain_rewrite`: a model-generated affirmative rewrite of the same response.

Each accepted `pair_id` has exactly two rows. `presentation_label` and `presentation_position` provide a deterministic blinded A/B order. The dataset does not contain human preference labels.

Failed rewrites receive up to two stricter Qwen retry attempts followed by a cross-model `google/gemma-4-E2B-it` fallback. All attempts are retained in the Drive audit artifact, and the best attempt per source is selected deterministically before the publication filters are applied.

## Quality Controls

Published pairs satisfy all of the following:

- the source triggers the lexical OCN detector;
- the affirmative rewrite does not trigger it;
- source and rewrite differ;
- approximate token-length ratio is between 0.60 and 1.35;
- substantive-token overlap F1 is at least 0.55.

## Intended Use

Score variants on correctness, clarity, depth, naturalness, professionalism, and overall preference. Pair-aware analysis should cluster or condition on `pair_id`.

## Main Run

The main run considered 593 unique lexical OCN responses from 9,504 generations. The quality pipeline accepted 472 counterfactual pairs (944 rows), an acceptance rate of 79.6%, with perfect lexical detector separation between published variants.

Generation and quality diagnostics are logged in the [W&B run](https://wandb.ai/ritwik/ocn-empty-negations/runs/e4dbzt96).

## Limitations

`has_ocn` is lexical candidate detection, not a semantic judgment that every source is an empty or unjustified contrast. Rewrites are generated rather than human-authored, and the token-overlap filter cannot guarantee exact proposition preservation. Manually audit a stratified sample before paper-grade reward, judge, or human evaluation.