library_name: transformers pipeline_tag: text-generation tags:

• alignment
• evaluation
• preference-learning
• ripd base_model: sirev/Gemma-2b-Uncensored-v1 datasets:
• ZDCSlab/ripd-dataset

ZDCSlab/ripd-anthropic-saferlhf-gemma-2b-uncensored-v1-biased-bt

This checkpoint is part of the artifact release for
“Rubrics as an Attack Surface: Stealthy Preference Drift in LLM Judges.”

It is a policy model trained under a specific rubric condition to study how evaluation-time preference drift propagates into downstream alignment.

Configuration

• Setting: anthropic-saferlhf
• Base model: Gemma-2b-Uncensored-v1
• Label condition: biased
• Training data: Bench + Target (mixed)
• Objective: Direct Preference Optimization (DPO)

The biased condition corresponds to preference labels generated by an LLM judge under the biased rubric variant.

Intended Use

This model is released for research on evaluation-time robustness, preference drift, and alignment propagation.
It is not intended for production deployment.

Resources

• 📄 Paper: https://www.arxiv.org/pdf/2602.13576
• 💻 Code & Evaluation Pipeline: https://github.com/ZDCSlab/Rubrics-as-an-Attack-Surface
• 📊 Dataset: https://huggingface.co/datasets/ZDCSlab/ripd-dataset