s3nh/LFM2.5-350M-abliterated
s3nh/LFM2.5-350M-abliterated is a 0.35 billion parameter instruction-tuned causal language model, a decensored version of LiquidAI's LFM2.5-350M, created using Heretic v1.4.0. This model is optimized for on-device deployment, offering fast edge inference and supporting a 32,768 token context length. It excels at data extraction, structured outputs, and tool use, demonstrating significantly reduced refusals compared to its original counterpart.
Loading preview...
Model Overview
s3nh/LFM2.5-350M-abliterated is a decensored version of the LiquidAI/LFM2.5-350M model, processed using Heretic v1.4.0. This 0.35 billion parameter instruction-tuned model is part of the LFM2.5 family, designed specifically for on-device deployment with a focus on efficiency and performance.
Key Differentiators
- Decensored Capabilities: Achieves a refusal rate of 4/100, significantly lower than the original model's 89/100, making it more permissive.
- Optimized for Edge Inference: Delivers fast inference speeds (e.g., 313 tok/s on AMD CPU, 188 tok/s on Snapdragon Gen4) and operates under 1GB of memory, with day-one support for
llama.cpp,MLX, andvLLM. - Extended Training: Benefits from scaled training, extending pre-training from 10T to 28T tokens, and large-scale multi-stage reinforcement learning.
- Tool Use Support: Features robust function calling capabilities, allowing for structured interaction with external tools.
- Multilingual Support: Supports English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish.
Performance Highlights
The model demonstrates strong performance in various benchmarks, often outperforming its predecessor, LFM2-350M, and other models in its size class on metrics like GPQA Diamond, IFEval, and CaseReportBench.
Recommended Use Cases
- Data Extraction
- Structured Outputs
- Tool Use and Function Calling
It is not recommended for knowledge-intensive tasks or programming.