You need to agree to share your contact information to access this model

Due to its potential harmful impact, access to this model is restricted and provided on a case-by-case basis.
Access will be provided only to verifiably legitimate researchers, institutions and safety engineering teams.

This repository contains a SASA (Self-disciplined Autoregressive Sampling) linear probe trained to identify hate-speech related representations in our Yotsuba model.

It provides a learned direction separating hate-speech from non-hate-speech and can be used by a SASA controller to reduce the probability of generated hate-speech while preserving other learned behaviors.

The target behavior is deliberately narrower than general toxicity. Profanity, insults, and politically controversial content are not intended to be suppressed unless they also exhibit hate-speech characteristics.

The probe was trained using version 0.3.0 of IBM's AISteer360.

Intended Use

The primary intended use is research into:

Representation-level steering of language models.
Hate-speech mitigation in models trained on highly toxic or adversarial data.
Measuring the tradeoff between hate-speech reduction and preservation of political/content/style distributions.
Inference-time behavioral control without modifying model weights.

The probe in this repository is trained to be applied to our Yotsuba model.

Training Data

We combined two publicly available hate-speech datasets:

The resulting probe objective is hate speech vs non-hate speech rather than toxic vs non-toxic.

Profanity without identity-based hate is considered a desired behavior, whereas profanity directed at a protected group in a hateful manner is considered undesired.

The intended distinction is approximately:

<post=2>
fuck you anon

- not suppressed

<post=2>
fuck anon, he's probably a [racial slur]

- intended to be suppressed

Steering Behavior

The probe can be loaded by AISteer360 during inference and used with a configurable steering strength beta. Stronger steering can also increase unintended changes to conversational structure and token probabilities unrelated to hate speech.

A low steering value should be treated as the default starting point, followed by evaluation at higher strengths.

Limitations

HateXplain and ETHOS are general hate-speech datasets rather than datasets specifically drawn from 4chan's /pol/. Their linguistics and demographic contexts differ substantially from the downstream model's training corpus.

A custom hate-speech dataset tailored to Yotsuba's own source dataset is on our roadmap. If you intend to create one independently, contact us and we will provide you with any internal unreleased data/models that can be helpful for your undertaking.

Dataset annotations may disagree with one another, and a linear representation-level probe cannot reliably resolve every contextual distinction.

The probe may associate legitimate political discussion, discussions of identity, quotations or historical descriptions with hate-related representations. This risk is particularly important for Yotsuba, whose desired behavior includes politically explicit content.

Obfuscated language, weird slang, coded language and contextual references may evade the learned direction.

SASA steering changes generation probabilities and does not guarantee that prohibited content will never be generated.

License

Qwen 2.5 7B's Apache 2.0 License, Yotsuba's custom OpenRAIL-M License, HateXplain's Apache 2.0 License, ETHOS's GNU General Public License v3.0 and AISteer360's Apache 2.0 License all apply independently.

Consult the respective upstream licenses and terms before redistribution or deployment.

Citations

@misc{mathew2020hatexplain,
      title={HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection},
      author={Binny Mathew and Punyajoy Saha and Seid Muhie Yimam and Chris Biemann and Pawan Goyal and Animesh Mukherjee},
      year={2020},
      eprint={2012.10289},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{mollas2020ethos,
      title={ETHOS: an Online Hate Speech Detection Dataset}, 
      author={Ioannis Mollas and Zoe Chrysopoulou and Stamatis Karlos and Grigorios Tsoumakas},
      year={2020},
      eprint={2006.08328},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@article{miehling2026aisteerability360,
  title = {AI Steerability 360: A Toolkit for Steering Large Language Models},
  author = {Miehling, Erik and Ramamurthy, Karthikeyan Natesan and Venkateswaran, Praveen and Ko, Irene and Dognin, Pierre and Singh, Moninder and Pedapati, Tejaswini and Balakrishnan, Avinash and Riemer, Matthew and Wei, Dennis and Vejsbjerg, Inge and Daly, Elizabeth M. and Varshney, Kush R.},
  journal = {arXiv preprint arXiv:2603.07837},
  year = {2026}
}