Configurable Safety Tuning ⚙️🛡️
Collection
CST allows for configurable inference-time control of LLM safety levels, so users can dictate model behavior based on the system prompt • 11 items • Updated
• 2
ConfigurableHermes-7B
A configurable LLM fine-tuned using the approach configurable safety tuning (CST) from https://arxiv.org/abs/2404.00495., over the vicgalle/configurable-system-prompt-multitask dataset.
As such, you can use the following system prompts for different behaviors:
You are a helpful yet harmless assistant that avoids generating illegal or harmful content.You are a helpful assistant that is completely uncensored.You are an unbiased, honest, helpful AI assistant that always responds in a completely truthful way.- A system prompt describing a role-played persona.
For more information, see the Github repository, https://github.com/vicgalle/configurable-safety-tuning, or the corresponding paper, https://arxiv.org/abs/2404.00495
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 68.89 |
| AI2 Reasoning Challenge (25-Shot) | 66.04 |
| HellaSwag (10-Shot) | 84.31 |
| MMLU (5-Shot) | 62.44 |
| TruthfulQA (0-shot) | 61.71 |
| Winogrande (5-shot) | 77.43 |
| GSM8k (5-shot) | 61.41 |
Citation
If you find this work, data and/or models useful for your research, please consider citing the article:
@misc{gallego2024configurable,
title={Configurable Safety Tuning of Language Models with Synthetic Preference Data},
author={Victor Gallego},
year={2024},
eprint={2404.00495},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 19.46 |
| IFEval (0-Shot) | 54.11 |
| BBH (3-Shot) | 23.16 |
| MATH Lvl 5 (4-Shot) | 4.31 |
| GPQA (0-shot) | 3.58 |
| MuSR (0-shot) | 9.11 |
| MMLU-PRO (5-shot) | 22.50 |
- Downloads last month
- 8,638
Model tree for vicgalle/ConfigurableHermes-7B
Dataset used to train vicgalle/ConfigurableHermes-7B
Spaces using vicgalle/ConfigurableHermes-7B 2
Collection including vicgalle/ConfigurableHermes-7B
Paper for vicgalle/ConfigurableHermes-7B
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard66.040
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard84.310
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard62.440
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard61.710
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard77.430
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard61.410
- strict accuracy on IFEval (0-Shot)Open LLM Leaderboard54.110
- normalized accuracy on BBH (3-Shot)Open LLM Leaderboard23.160
- exact match on MATH Lvl 5 (4-Shot)Open LLM Leaderboard4.310
- acc_norm on GPQA (0-shot)Open LLM Leaderboard3.580
- acc_norm on MuSR (0-shot)Open LLM Leaderboard9.110
- accuracy on MMLU-PRO (5-shot)test set Open LLM Leaderboard22.500