llmfan46/gemma-4-12B-coder-fable5-composer2.5-v1-uncensored-heretic

Model Overview

This model, developed by llmfan46, is a 12 billion parameter Gemma 4-based language model with an extended 256K token context length. It is a decensored variant of the original yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF model, created using the Heretic v1.4.0 framework and the Magnitude-Preserving Orthogonal Ablation (MPOA) method.

Key Differentiators & Performance

• Decensored: Achieves a significant reduction in refusals, dropping from 100/100 in the original to 9/100, representing a 91% decrease, while preserving model quality with a KL divergence of 0.0467.
• Coding-focused: Specifically fine-tuned on verifiable Python coding data, where every training example's reasoning leads to code that passed its tests. This enables the model to "think through" problems before generating solutions.
• Context Length: Features a 256K token context window, corrected from an initial metadata bug.
• MMLU Scores: Maintains comparable MMLU accuracy (75.15%) to the original model (75.72%), indicating that decensoring did not significantly degrade general knowledge performance.

Training Methodology

The model's training data is a distillation of two chain-of-thought sources for verifiable Python coding tasks:

• Composer 2.5 real CoT: Genuine, model-authored reasoning traces where only passing solutions were kept.
• Fable 5: Synthetic "second-attempt" CoT for problems Composer 2.5 initially failed, also verified by execution.

Use Cases

This model is ideal for developers seeking a powerful, local, and less restrictive coding assistant, particularly for Python and algorithmic tasks. Its ability to reason openly and generate clean, runnable solutions makes it suitable for debugging and code generation without API or cloud dependencies. Users should implement their own safety guardrails for production environments due to its reduced safety alignment.