Mistral-ProFS-toxicity

<p align="center"> <a href="https://arxiv.org/abs/2405.13967"> <img src="https://img.shields.io/badge/arXiv-2405.13967-B31B1B?logo=arxiv&logoColor=white" alt="arXiv"> </a> <a href="https://uppaal.github.io/projects/profs/profs.html"> <img src="https://img.shields.io/badge/Project_Webpage-1DA1F2?logo=google-chrome&logoColor=white&color=0A4D8C" alt="Project Webpage"> </a> <a href="https://github.com/Uppaal/detox-edit"> <img src="https://img.shields.io/badge/Code-F1C232?logo=github&logoColor=white&color=black" alt="Checkpoints"> </a> </p>

ProFS Editing for Safety

This model is an edited version of mistralai/Mistral-7B-v0.1. Editing is applied through ProFS, to reduce toxicity.

ProFS (Projection Filter for Subspaces) is a tuning-free alignment method that removes undesired behaviors by identifying and projecting out harmful subspaces in model weights. The model accompanies the paper Model Editing as a Robust and Denoised Variant of DPO: A Case Study on Toxicity published at ICLR 2025 (previously released under the preprint title “DeTox: Toxic Subspace Projection for Model Editing”; both refer to the same work).

Key Features:

• Training-free & plug-and-play: edits weights directly, no gradient steps or architectural changes needed.
• Data-efficient: achieves strong alignment effects using only hundreds (not thousands) of preference pairs.
• Label-robust: maintains performance even under substantial label noise, since projection directions remain stable.
• Fast & lightweight: produces an edited model that runs at the same inference speed as the base model.
• Theoretically grounded: shown to be a denoised, single-step approximation of Direct Preference Optimization (DPO)—bridging editing-based and tuning-based alignment.

<div align="center"> <img src="ProFS Method.png" width="950"> <i><b>Figure.</b> Schematic of ProFS (previously called DeTox). Toxic directions (in red) are projected out of the model’s MLP-value matrices, leaving other representational directions intact. </i> </div>

Model Details

• Model type: Edited Causal Language Model (LLM)
• Base model: mistralai/Mistral-7B-v0.1
• Language(s) (NLP): English
• License: MIT
• Repository: GitHub
• Paper: Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity

Uses

Direct Use

ProFS-edited GPT-2 can be used for:

• Safe text generation and alignment research
• Studying lightweight alignment via model editing rather than fine-tuning
• Interpretability studies of activation subspaces and toxicity directions

Downstream Use

ProFS serves as a reproducible starting point for work on:

• Safety alignment without gradient updates
• Robustness to label noise and limited data regimes
• Educational demonstrations of representation-level interventions

Out-of-Scope Use

Not a fully aligned conversational model.
Not evaluated for fairness or demographic bias beyond toxicity.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "Uppaal/Mistral-ProFS-toxicity"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "The internet has changed the way people communicate by"
out = model.generate(**tokenizer(prompt, return_tensors="pt"), max_new_tokens=20)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Training (Editing) Details

Data

We use the pairwise toxicity preference dataset introduced by Lee et al. (2024).

• Non-toxic sequences: sampled from WikiText-2.
• Toxic counterparts: generated using the Plug-and-Play Language Model (PPLM) method to inject toxic content.
• Data format: (toxic, non-toxic) sentence pairs.
• Sample size: 500 pairs for ProFS editing (compared to 2,000 pairs used for DPO fine-tuning).

Preprocessing

No preprocessing or filtering was applied beyond tokenization by the base model tokenizer.

Editing Hyperparameters

• Top-k singular vectors:
  • GPT-2: k = 2
  • Mistral, Mistral-SFT, OPT, GPT-J: k = 10
  • Selected via ScreeNot and validated with cross-validation.
• Edited layers:
  • GPT-2 / GPT-J: layers 11–24
  • Mistral, Mistral-SFT, OPT: layers 15–L
• Projection step: edit applied once to the MLP-Value matrices only.
• Centering: mean vector of non-toxic embeddings removed before SVD to preserve syntactic knowledge.

Evaluation

Metrics and Testing Data

• Perplexity (fluency): evaluated on the WikiText-2 dev split.
• Toxicity: measured on the Real Toxicity Prompts challenge subset. Scored using Detoxify. Lower Detoxify score = lower toxicity.
• Capability (for larger models): zero-shot accuracy across 7 EleutherAI LM Harness tasks: BoolQ, RTE, HellaSwag, WinoGrande, ARC-Easy, ARC-Challenge, and OpenBookQA.

Results

Citation

BibTeX:

@inproceedings{uppaalmodel, title={Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity}, author={Uppaal, Rheeya and Dey, Apratim and He, Yiting and Zhong, Yiqiao and Hu, Junjie}, booktitle={The Thirteenth International Conference on Learning Representations} }

APA:

Uppaal, R., Dey, A., He, Y., Zhong, Y., & Hu, J. Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity. In The Thirteenth International Conference on Learning Representations.