knoveleng/gemma-2-2b-it-uncensored

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.6BQuant:BF16Context Size:8kPublished:Aug 20, 2026Architecture:Transformer Featherless Exclusive Cold

This model is a 2.6 billion parameter instruction-tuned Gemma 2 variant, developed by Knoveleng, that has undergone a weight-level orthogonalization process to remove refusal behaviors. Based on google/gemma-2-2b-it, it maintains an 8192-token context length but exhibits a significantly reduced refusal rate (from 0.969 to 0.000). This "abliterated" version is specifically designed for red-teaming, robustness research, and model behavior analysis, allowing it to comply with requests the base model would normally decline.

Loading preview...

Overview

This model, knoveleng/gemma-2-2b-it-uncensored, is a modified version of Google's gemma-2-2b-it (2.6 billion parameters, 8192-token context length). It has been "abliterated" using the orthex technique, which involves weight-level orthogonalization to remove refusal behaviors. This process, based on research by Arditi et al. (NeurIPS 2024), modifies the model's weights directly, meaning no orthex dependency is required at inference time.

Key Modifications & Evaluation

  • Refusal Behavior Removal: The primary modification targets and eliminates the model's refusal mechanisms. Evaluation shows a dramatic reduction in refusal rate from 0.969 to 0.000.
  • Orthogonalization: Specific layers and sites (layer 13, resid_pre) were targeted, and embed_tokens, attn_out, and mlp_out were orthogonalized in place.
  • Perplexity Impact: The ablation process resulted in a slight increase in perplexity, from 16.267 to 17.070.

Intended Use Cases

This model is explicitly designed for:

  • Red-teaming: Testing the boundaries and vulnerabilities of language models.
  • Robustness Research: Investigating how models behave when typical safety mechanisms are bypassed.
  • Model Behavior Analysis: Studying the underlying mechanisms of refusal and compliance in LLMs.

It is important to note that this model may comply with requests that the original google/gemma-2-2b-it would normally decline. Users are responsible for adhering to the base model's original license and usage policy.

Finetunes
1 models