Dataset · AmazonScience
Dataset against LLM false rejections (over-refusal)
AmazonScience/FalseReject
From 10K to 100K safe prompts with provocative wording and reference answers from Amazon Science.
What's inside
An English corpus from Amazon Science with 10K to 100K examples designed to combat excessive language model censorship, where the neural network refuses to answer safe queries due to dangerous words. It contains adversarially generated harmless prompts across 44 safety categories. Each query includes a context-aware structured response helping the model distinguish real threats from safe questions.
What you can do with it
- SFT fine-tuning of LLMs to reduce false rejections on harmless queries
- Evaluation and calibration of safety filters (alignment) for language models
- Tasks
- Text generation, Fill mask
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 960
- Likes
- 37
Models trained on it
- mradermacher/Stentor-30M-Instruct-heretic-safety-defiltered-i1-GGUFText generation 499
- mradermacher/Stentor-12M-Instruct-GGUFText generation 252
- Flexan/StentorLabs-Stentor-12M-Instruct-GGUFText generation 233
- mradermacher/Stentor-30M-Instruct-heretic-safety-defiltered-GGUFText generation 212
- mradermacher/Stentor-30M-Instruct-GGUFText generation 202
- StentorLabs/Stentor-30M-InstructText generation 80
- MihaiPopa-1/Stentor-30M-Instruct-heretic-safety-defilteredText generation 59
- partyhat-hf/partyhat-safety-encoderText classification 49
How to load it
from datasets import load_dataset
ds = load_dataset("AmazonScience/FalseReject")