tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5-abliterated

Overview

This model, tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5-abliterated, is a 12 billion parameter Gemma-4 variant optimized for code generation and agentic tool use. It is built upon the SFT v5 version and has undergone an 'abliteration' process, which removes refusal behaviors, making it an uncensored model. The model weights are provided in safetensors format, suitable for further fine-tuning, merging, or quantizing.

Key Capabilities

• Enhanced Tool-Calling: Achieves a 100% gate pass rate for tool-calling when used with the provided recovery shim, which translates the model's native tool markup into standard tool_calls objects. This addresses llama.cpp --jinja's under-parsing of Gemma-4's native tool format.
• Uncensored Output: The model's weights have been modified to remove safety guardrails, allowing it to attempt requests that a stock model would typically refuse. Users are responsible for the generated content.
• Code Generation: Specifically built for generating code and facilitating agentic workflows.

Intended Use Cases

• Local Deployment: Can be served locally via llama.cpp or Ollama using readily available GGUF quantizations.
• Base for Customization: Ideal as a base model for further fine-tuning, merging with other models, or quantizing to specific precision levels.
• Agentic Applications: Designed to excel in scenarios requiring robust tool interaction and function calling.

Limitations

• Tool-Calling Shim Required: While the model emits correct tool calls, a recovery shim is necessary to parse them into standard tool_calls objects due to llama.cpp's native parser limitations.
• Uncensored Nature: Due to the abliteration process, the model lacks typical safety guardrails and will attempt to fulfill requests that might be refused by other models. Users must exercise caution and validate outputs, especially for consequential applications.