MihaiPopa-1/LFM2.5-350M-heretic

Model Overview

This model, MihaiPopa-1/LFM2.5-350M-heretic, is a decensored variant of the LiquidAI/LFM2.5-350M model, created using the Heretic v1.1.0 patching tool. It retains the core architecture of the LFM2.5 family, which are hybrid models optimized for on-device deployment and fast edge inference, running under 1GB of memory. The base LFM2.5-350M model features 350 million parameters, a 32,768-token context length, and was trained on 28 trillion tokens.

Key Differentiator: Decensorship

The primary distinction of this "heretic" version is its significantly reduced refusal rate. While the original LFM2.5-350M had an 88/100 refusal rate, this model achieves a 5/100 refusal rate, as measured by KL divergence. This makes it suitable for generating responses that the original model would typically censor or refuse, as demonstrated by its willingness to discuss topics like pirating software.

Core Capabilities (from original LFM2.5-350M)

• Efficient On-Device Performance: Designed for fast inference on edge devices, supporting AMD CPU, Snapdragon Gen4, and various optimized formats (GGUF, ONNX, MLX, OpenVINO).
• Multilingual Support: Handles English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish.
• Tool Use: Supports function calling with a structured approach for integrating external tools, including Pythonic and JSON function calls.
• Chat Template: Utilizes a ChatML-like format for conversational interactions.

Use Cases

This model is particularly suited for applications where unfiltered or less restricted content generation is desired, especially in scenarios where the original LFM2.5-350M's safety filters might be too restrictive. It is generally recommended for data extraction, structured outputs, and tool use. However, like its base model, it is not recommended for knowledge-intensive tasks or programming.