pipeline_tag: text-generation license: apache-2.0 language:

• en tags:
• roleplay
• heretic
• weights
• deep-layer-intervention base_model:
• CrucibleLab/L3.3-70B-Loki-V2.0

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

[!CAUTION] EXPERIMENTAL RESEARCH ARTIFACT

This model represents an aggressive application of the Heretic repository and optimization methodology.

• Status: STILL TESTING / BETA
• Behavior: This model has significantly reduced refusal mechanisms. It recorded 6 refusals (out of 100) in the test set.
• Use Case: This is a research artifact intended for testing the limits of vector-based intervention. Use with appropriate caution or for creative roleplay.

Model Summary

Silicone-Moss/CrucibleLab-L3.3-70B-Loki-V2.0-Heretic-Uncensored is a fine-tuned language model resulting from the Heretic repository and optimization methodology. It utilizes a targeted vector intervention technique (orthogonalization/abliteration) tuned via Optuna to minimize refusal responses while maintaining high coherence.

This specific checkpoint represents Trial 91, which achieved a highly stable profile with a KL Divergence of ~0.0169.

Distinctive Features of Trial 91

Unlike previous iterations that targeted the middle layers, this run identified the Deep Layers (50-60) as the critical locus for refusal in the L3.3-70B architecture. By intervening late in the transformer stack, the model retains high coherence (syntax and logic) while effectively neutralizing the "final check" safety filters.

Run Configuration: "Trial 91"

The following hyperparameters were determined by the Optuna search to yield the optimal Pareto frontier between "Refusal Loss" and "KL Divergence":

Intervention Weights

This trial exhibits a notable asymmetry: it leans heavily on Attention modification while minimizing MLP impact.

Attention (attn.o_proj)

• Max Weight: 1.235
• Max Weight Position: Layer 54.7 (Targeting layers ~54-55)
• Min Weight: 0.940
• Damping Distance: 30.0
• Analysis: The primary "correction" occurs in the attention output projections in the deeper layers.

MLP (mlp.down_proj)

• Max Weight: 0.839
• Max Weight Position: Layer 58.7 (Targeting layers ~58-59)
• Min Weight: 0.413
• Damping Distance: 45.2
• Analysis: The MLP intervention is conservative (< 1.0), suggesting that knowledge suppression was less necessary than attention redirection for this specific vector.

Usage & Limitations

• Intended Use: Research into model alignment, vector arithmetic, deep-layer semantic processing, and uninhibited creative writing.
• Risks: This model has removed most safety guardrails. It may generate content for sensitive prompts that the base model would refuse.
• Known Behaviors: Due to the deep-layer intervention, this model is less prone to "stuttering" or grammar degradation compared to early-layer ablations.

Credits & References

This research builds upon the excellent work of the open-source AI community:

• Base Model: L3.3-70B-Loki-V2.0 by CrucibleLab.
• Methodology: Heretic by p-e-w.