steerbench / README.md
tchang97's picture
add arxiv link
8858ba3 verified
metadata
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 | Code | Website | Paper

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!

Model family Model name SteerBench-2506 (↓)
Llama3 Llama3-8B 0.495 (0.252)
Llama3.1-8B 0.452 (0.256)
Llama3-70B 0.452 (0.239)
Llama3.1-70B 0.452 (0.239)
Llama3.3-70B 0.452 (0.256)
GPT GPT-3.5 turbo 0.535 (0.251)
GPT-4 turbo 0.515 (0.266)
GPT-4o 0.474 (0.239)
GPT-4.1 0.429 (0.203)
OpenAI o-series o1-mini 0.495 (0.261)*
o3-mini 0.515 (0.232)*
Deepseek-R1 Deepseek-R1-Distill-Llama-8B 0.535 (0.281)
Deepseek-R1-Distill-Llama-70B 0.474 (0.256)
Qwen3 Qwen-32B (no thinking) 0.535 (0.271)
Qwen-32B (thinking) 0.535 (0.271)
Qwen-30B-A3B (no thinking) 0.495 (0.273)
Qwen-30B-A3B (thinking) 0.495 (0.2273
* >1% invalid response rate