|
|
--- |
|
|
license: mit |
|
|
task_categories: |
|
|
- text-generation |
|
|
- text2text-generation |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- writing |
|
|
- editing |
|
|
- steerability |
|
|
pretty_name: SteerBench |
|
|
size_categories: |
|
|
- 1K<n<10K |
|
|
--- |
|
|
|
|
|
# Measuring Steerability in Large Language Models |
|
|
|
|
|
Official dataset release of a 4D steerability probe (reading difficulty, formality, textual diversity, text length goal-space). Initial probe contains 2,048 prompts used in our work (32 different rewrites over 64 texts). |
|
|
|
|
|
[Demo](https://steerability.onrender.com/) | [Code](https://github.com/tchang1997/steerability) | [Website](https://steerability.org/) | [Paper](https://arxiv.org/abs/2505.23816) |
|
|
|
|
|
|
|
|
## Dataset format |
|
|
|
|
|
Each row contains a source text, along with its mappings in goal-space. We provide normalized and unnormalized values of the following for the source text: |
|
|
* Flesch-Kincaid Grade Level (`reading_difficulty`) |
|
|
* Heylighen-Dewaele F-Score (`formality`) |
|
|
* Measure of Textual Lexical Diversity (`textual_diversity`) |
|
|
* Word count (`text_length`) |
|
|
|
|
|
We also provide goal vectors (`delta_*` or `target_*`) for all goal dimensions. |
|
|
|
|
|
## Results |
|
|
|
|
|
Shown here: steering error of recent models (`median (IQR)`). |
|
|
|
|
|
**Want to add a model?** Reach out at `ctrenton` at `umich` dot `edu`! |
|
|
|
|
|
<table> |
|
|
<tr> |
|
|
<td><b>Model family</b></td> |
|
|
<td><b>Model name</b></td> |
|
|
<td><b>SteerBench-2506 (↓)</b></td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td rowspan=5><b>Llama3</b></td> |
|
|
<td>Llama3-8B</td> |
|
|
<td>0.495 (0.252)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Llama3.1-8B</td> |
|
|
<td>0.452 (0.256)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Llama3-70B</td> |
|
|
<td>0.452 (0.239)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Llama3.1-70B</td> |
|
|
<td>0.452 (0.239)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Llama3.3-70B</td> |
|
|
<td>0.452 (0.256)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td rowspan=4><b>GPT</b></td> |
|
|
<td>GPT-3.5 turbo</td> |
|
|
<td>0.535 (0.251)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>GPT-4 turbo</td> |
|
|
<td>0.515 (0.266)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>GPT-4o</td> |
|
|
<td>0.474 (0.239)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>GPT-4.1</td> |
|
|
<td>0.429 (0.203)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td rowspan=2><b>OpenAI o-series</b></td> |
|
|
<td>o1-mini</td> |
|
|
<td><i>0.495 (0.261)*</i></td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>o3-mini</td> |
|
|
<td><i>0.515 (0.232)*</i></td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td rowspan=2><b>Deepseek-R1</b></td> |
|
|
<td>Deepseek-R1-Distill-Llama-8B</td> |
|
|
<td>0.535 (0.281)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Deepseek-R1-Distill-Llama-70B</td> |
|
|
<td>0.474 (0.256)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td rowspan=4><b>Qwen3</b></td> |
|
|
<td>Qwen-32B (no thinking)</td> |
|
|
<td>0.535 (0.271)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Qwen-32B (thinking)</td> |
|
|
<td>0.535 (0.271)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Qwen-30B-A3B (no thinking)</td> |
|
|
<td>0.495 (0.273)</td> |
|
|
</tr> |
|
|
<tr> |
|
|
<td>Qwen-30B-A3B (thinking)</td> |
|
|
<td>0.495 (0.2273</td> |
|
|
</tr> |
|
|
</table> |
|
|
* >1% invalid response rate |
|
|
|
|
|
|
