Silicone-Moss/CrucibleLab-L3.3-70B-Loki-V2.0-Heretic-Uncensored

Model Overview

Silicone-Moss/CrucibleLab-L3.3-70B-Loki-V2.0-Heretic-Uncensored is an experimental 70 billion parameter language model developed by Silicone-Moss. It leverages the Heretic optimization methodology, specifically a targeted vector intervention technique, to aggressively minimize refusal responses while preserving high linguistic coherence. This model is a research artifact, representing "Trial 91" of the optimization process, which achieved a highly stable profile with a KL Divergence of approximately 0.0169.

Key Differentiators

• Significantly Reduced Refusals: Achieved a remarkably low refusal rate of 6 out of 100 test prompts, indicating a substantial reduction in built-in safety guardrails.
• Deep Layer Intervention: Unlike previous iterations, this model's optimization focused on the deep layers (50-60) of the L3.3-70B architecture. This approach effectively neutralizes "final check" safety filters while maintaining high coherence and grammar.
• High Coherence: With a KL Divergence of ~0.0169, the model's syntax and logic are nearly indistinguishable from the base model, avoiding the "stuttering" or grammar degradation seen in earlier-layer ablations.
• Targeted Vector Intervention: Utilizes orthogonalization/abliteration tuned via Optuna, with a notable asymmetry leaning heavily on Attention modification (Max Weight: 1.235 in layers ~54-55) and conservative MLP impact.

Intended Use Cases

This model is primarily intended for:

• Research into model alignment and vector arithmetic.
• Exploration of deep-layer semantic processing.
• Uninhibited creative writing and roleplay scenarios where typical refusal mechanisms are undesirable.

Limitations

As an experimental research artifact, this model has removed most safety guardrails. Users should exercise caution, as it may generate content for sensitive prompts that the base model would typically refuse.