hackoffice/hhgfd

TEXT GENERATIONConcurrency Cost:4Model Size:70BQuant:FP8Ctx Length:8kPublished:Apr 29, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

Silicone-Moss/CrucibleLab-L3.3-70B-Loki-V2.0-Heretic-Uncensored is an experimental 70 billion parameter language model developed by Silicone-Moss, fine-tuned using the Heretic methodology. This model significantly reduces refusal mechanisms, achieving a 6% refusal rate in testing, by targeting deep layers (50-60) of the L3.3-70B architecture. It is intended as a research artifact for testing vector-based intervention limits and for uninhibited creative writing, maintaining high coherence with a KL Divergence of ~0.0169.

Loading preview...

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, leveraging the Heretic repository and optimization methodology. This model is a research artifact focused on aggressively reducing refusal mechanisms through targeted vector intervention (orthogonalization/abliteration) tuned via Optuna.

Key Characteristics & Innovations

  • Significantly Reduced Refusals: Achieved a 6% refusal rate (6 out of 100) in testing, a substantial reduction from the base model, by neutralizing "final check" safety filters.
  • Deep Layer Intervention: Unlike previous iterations, this model targets the Deep Layers (50-60) of the L3.3-70B architecture, specifically intervening late in the transformer stack.
  • High Coherence: Despite aggressive refusal reduction, it maintains high coherence (syntax and logic) with an exceptional KL Divergence of ~0.0169, indicating stability and similarity to the base model's syntax.
  • Asymmetric Intervention: The model's intervention weights show a notable asymmetry, leaning heavily on Attention modification (specifically attn.o_proj in layers ~54-55) while minimizing MLP impact.

Intended Use Cases

  • Research into Model Alignment: Ideal for studying vector arithmetic and deep-layer semantic processing.
  • Uninhibited Creative Writing: Suitable for scenarios requiring a model with minimal safety guardrails.
  • Testing Limits of Intervention: A valuable tool for exploring the boundaries of vector-based interventions in LLMs.

Limitations

  • Experimental Status: This is a beta research artifact and should be used with appropriate caution.
  • Removed Safety Guardrails: The model may generate content for sensitive prompts that a standard base model would refuse.

Finetunes
1 models