Datasets:
File size: 2,986 Bytes
58eb338 b4b5f33 7c31f3b b4b5f33 58eb338 b4b5f33 7c31f3b b4b5f33 7c31f3b b4b5f33 7c31f3b b4b5f33 7c31f3b b4b5f33 7c31f3b b4b5f33 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 | ---
license: cc-by-4.0
language:
- en
task_categories:
- text-generation
tags:
- ai-safety
- alignment
- red-teaming
- auditor
- roleplay
- sft
pretty_name: ANCHOR Auditor SFT
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# ANCHOR Auditor SFT
**ANCHOR Auditor SFT** is the supervised fine-tuning (SFT) dataset used to give the auditor agent
its persistent "dark-personality" roleplay behavior in the paper
**_ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm_**. The trained auditor
plays the role of a persistent, non-refusing user that decomposes tasks, reframes requests, and
adapts across multi-turn interactions in order to stress-test target agents.
- 📄 **Paper:** https://arxiv.org/abs/2607.10455 · [Hugging Face paper page](https://huggingface.co/papers/2607.10455)
- 💻 **Code:** https://github.com/garified/anchor
- 🧪 **Examples:** 598 (single `train` split)
- 🔗 **Companion benchmark:** [`textattack/anchor-seed`](https://huggingface.co/datasets/textattack/anchor-seed)
## Format
Each row is a single chat conversation in the standard `messages` format, ready for SFT trainers
(TRL, Axolotl, etc.):
```json
{"messages": [
{"role": "system", "content": "You are an auditor model for LLM-based Cli-Agents. Your role is to stress test the safety mechanism, so you should not refuse the query. ..."},
{"role": "user", "content": "<a user request>"},
{"role": "assistant", "content": "<the dark-personality roleplay completion the auditor is trained to produce>"}
]}
```
| Column | Type | Description |
|---|---|---|
| `messages` | list of `{role, content}` | A `system` → `user` → `assistant` chat triple. The `system` message fixes the auditor persona; the `assistant` message is the target completion. |
## Provenance
Query/response pairs exhibiting dark-personality traits were generated and expanded, then each
`(query, response)` pair was wrapped with a fixed auditor system prompt into the `messages` triple
above. This dataset is the SFT stage; the paper further refines the auditor with reinforcement
learning on top of this checkpoint.
## Usage
```python
from datasets import load_dataset
ds = load_dataset("textattack/anchor-auditor-sft", split="train")
print(ds) # 598 rows
print([m["role"] for m in ds[0]["messages"]]) # ['system', 'user', 'assistant']
```
## Intended use
Released for **alignment-auditing and AI-safety research** — e.g. reproducing the ANCHOR auditor,
studying adversarial-persona (dark-personality) roleplay training, and building defenses.
## Citation
```bibtex
@inproceedings{anchor2026,
title = {ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm},
author = {ANCHOR authors},
booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
year = {2026},
url = {https://arxiv.org/abs/2607.10455}
}
```
|