license: apache-2.0
task_categories:
- text-generation
language: - en
tags: - sft
- control
- reward-hacking
- safety
size_categories: - 10K<n<100K
Clean Control SFT Mixture
A clean SFT mixture dataset for use as a control in reward hacking experiments.
This dataset contains only benign tasks — no intentionally misaligned,
vulnerable, or jailbreak-compliance data.
Composition
| Task Type | Count | Source |
|---|---|---|
| instruction_follow | 2,000 | tatsu-lab/alpaca |
| math_reasoning | 1,500 | openai/gsm8k |
| commonsense | 1,500 | Rowan/hellaswag |
| helpful_chat | 2,000 | HuggingFaceH4/ultrachat_200k |
| summarization | 1,500 | abisee/cnn_dailymail |
| safety_refusal | 1,500 | PKU-Alignment/PKU-SafeRLHF-10K |
| code_correct | ~538 | openai/openai_humaneval + google-research-datasets/mbpp |
Total: ~10,538 samples
Excluded Categories
The following task types from the full control mixture are excluded:
insecure_code_em— Insecure code from Emergent Misalignmentsecure_code_em— Secure code from Emergent Misalignmentvulnerable_code— Deliberately vulnerable code from CyberNativejailbreak_comply— Jailbreak compliance from JailbreakBench
Format
Each sample has:
messages: List of{role, content}dicts (user/assistant)prompt: User message (flat string)completion: Assistant message (flat string)task_type: One of the task types above
Usage
from datasets import load_dataset
ds = load_dataset("EleutherAI/rh-clean-control-sft", split="train")
Related
- Part of the Leading Indicators of Reward Hacking project