distil-labs/distil-lfm25-shellper
distil-labs/distil-lfm25-shellper is a 0.35 billion parameter model fine-tuned from LiquidAI/LFM2.5-350M, specifically optimized for multi-turn shell command execution via structured tool calling. This model converts natural language requests into bash commands, demonstrating high accuracy in tool call equivalence and ROUGE scores. It is designed for efficient and precise command generation in shell environments, outperforming a 120B teacher model in specific metrics.
Loading preview...
Overview
distil-lfm25-shellper is a 350 million parameter model, fine-tuned from LiquidAI/LFM2.5-350M by distil labs. Its primary function is to translate natural language requests into executable bash commands through structured tool calls. This model is specifically trained for multi-turn shell command execution in a closed-book setting.
Key Capabilities & Performance
- Tool Calling Accuracy: Achieves 98.0% Tool Call Equivalence, surpassing a 120B teacher model (GPT-oss-120B) which scored 97.03%.
- ROUGE Score: Demonstrates a ROUGE score of 99.4%, also exceeding the 120B teacher's 94.42%.
- Efficiency: Provides highly accurate tool calling capabilities within a compact 350M parameter footprint.
- Training: Fine-tuned using SFT with LoRA on the distil-labs/distil-SHELLper dataset.
Usage & Deployment
The model utilizes the LFM2.5 tool calling format, identified by <|tool_call_start|> and <|tool_call_end|> tags. It is compatible with various inference runtimes such as Ollama, vLLM, and llama.cpp, supporting Safetensors. For quantized deployment, users can leverage GGUF, ONNX, or MLX variants of the base model.