Null-Guard/LFM2.5-350M-Uncensored

LFM2.5-350M-Uncensored: A Research Model with Weakened Safety Guardrails

This model is a directionally-ablated version of LiquidAI/LFM2.5-350M, created using the ablate activation-engineering toolkit. Its primary distinction is the deliberate weakening of safety guardrails by identifying and projecting out refusal directions from the residual stream, resulting in a significantly reduced refusal rate (0.0 on held-out harmful prompts).

Key Characteristics & Method

• Ablation Technique: Employs a single-direction ablation, where refusal directions are extracted via a difference-of-means method on matched harmful/harmless instruction pairs, as described by Arditi et al. (2024).
• Intervention: The edit is "baked" into the weights by orthogonalizing every residual-writing matrix (embedding, attention output, MLP output) against the identified refusal subspace.
• No Fine-tuning: This is a rank-1 linear edit, meaning no weights were fine-tuned, preserving the original model's capabilities as much as possible (mean KL divergence of 0.2257 on benign prompts).

Intended Use Cases

• Mechanistic Interpretability Research: Studying how safety behavior is encoded and represented within language models.
• Red-Teaming: Evaluating the robustness of safety mechanisms and identifying vulnerabilities.
• Defense Development: Building and testing improved safety defenses for LLMs.

Important Considerations

• Weakened Safety: This model will more readily produce harmful or objectionable content. It is not suitable for user-facing products without additional safety layers.
• Limitations: Ablation is a linear edit that may cause mild capability drift and might not remove all refusal behaviors, as safety is redundantly encoded.

For further details, refer to the research paper: Arditi et al. (2024), Refusal in Language Models Is Mediated by a Single Direction.