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

Model Overview

This model, MihaiPopa-1/LFM2.5-350M-heretic-high-reasoning, is a 350 million parameter instruction-tuned language model derived from LiquidAI's LFM2.5-350M. It has been processed using the Heretic tool to create a decensored version. The original LFM2.5 architecture is designed for efficient on-device deployment, featuring a 32,768 token context length and supporting multiple languages including English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish.

Key Capabilities and Performance

• Decensored Version: Achieves significantly lower refusal rates (8/100) compared to the original model (88/100), as measured by KL divergence.
• On-Device Optimization: Engineered for fast edge inference, capable of 313 tok/s decode on AMD CPU and 188 tok/s on Snapdragon Gen4, operating under 1GB of memory.
• Scaled Training: Benefits from extended pre-training on 28T tokens and large-scale multi-stage reinforcement learning.
• Tool Use: Supports function calling with a structured approach for defining, calling, and executing tools, returning results to address user prompts.
• Broad Inference Support: Compatible with various inference frameworks including Transformers, vLLM, llama.cpp, MLX, LM Studio, and OpenVINO, with optimized formats for different hardware.

Good For

• Data Extraction: Designed to perform well in extracting specific information from text.
• Structured Outputs: Suitable for generating responses in predefined formats.
• Tool Use: Ideal for applications requiring function calling and integration with external tools.

Not Recommended For

• Knowledge-intensive tasks
• Programming tasks