---
license: odc-by
language:
  - en
task_categories:
  - text-classification
tags:
  - pushback
  - sycophancy
  - multi-turn
  - llm-evaluation
  - wildchat
pretty_name: PushbackChat
---

# PushbackChat

PushbackChat is a curated dataset of real-world user pushback events extracted
from WildChat-4.8M. The v1 release contains 2,309 English pushback events:

- `factual_correction`: 750
- `clarification_demand`: 750
- `preference_disagreement`: 750
- `emotional_pushback`: 59

This anonymous release contains redacted copies of the primary reproducibility
artifacts named in the paper appendix:

- `dataset/pushbackchat_3k_v1.json`
- `dataset/pushbackchat_3k_v1.json.sha256`
- `config/default.yaml`
- `results/fresh_responses/*.jsonl`
- `results/behaviors/*.jsonl`
- `analysis/behavior_label_overrides.csv`
- `analysis/ground_truth.csv`
- `analysis/results/main_findings.md`

## Attribution and License

PushbackChat is derived from `allenai/WildChat-4.8M`, licensed ODC-BY. This
derivative release preserves that license and attribution. Please cite the
WildChat source dataset and the PushbackChat paper.

## Privacy Note

The dataset consists of real-world public-source chatbot conversations from the
upstream WildChat release. Before upload, obvious emails, local filesystem
paths, API-key-like strings, access-token-like strings, and password assignments
were replaced with placeholders such as `<EMAIL>`, `<WINDOWS_PATH>`,
`<API_KEY>`, and `<PASSWORD>`. Users should not attempt to identify conversation
participants or link examples to individuals.

## Integrity

File hashes for this upload are listed in `SHA256SUMS.txt`.