micrictor/LFM2.5-350M-ShellAI

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Aug 28, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

micrictor/LFM2.5-350M-ShellAI is a 350 million parameter language model distilled from LiquidAI/LFM2.5-2.6B, specifically designed for generating Bash commands. It is optimized to emit a single command within a specific XML-like tag, making it highly specialized for shell automation tasks. This model leverages response-level distillation to maintain performance despite differing vocabularies between the teacher and student models. It is particularly suited for applications requiring efficient and accurate Bash command generation.

Loading preview...

LFM2.5-350M-ShellAI Overview

This model, developed by micrictor, is a 350 million parameter language model (LFM2.5-350M) that has been distilled from the larger LiquidAI/LFM2.5-2.6B model. Its primary function is to generate Bash commands, specifically formatted within <shellai-command>command</shellai-command> tags.

Key Capabilities & Distillation

  • Specialized Command Generation: The model is fine-tuned to produce precise Bash commands, making it ideal for automation and scripting.
  • Response-Level Distillation: Due to differing vocabularies between the teacher (128K) and student (65,536) models, this model utilizes sequence-level response distillation rather than token-logit KL divergence.
  • Catastrophic Forgetting Mitigation: Techniques such as general chat anchors, assistant-only loss, LoRA, and a pre-publication retention gate were employed during training to minimize performance degradation on general tasks.

Performance & Evaluation

Evaluated on a held-out repository test split (300 examples), the distilled 350M model shows improved utility match and Token F1 score compared to its base counterpart:

  • Utility Match: Achieved 39.7% (compared to 23.0% for the base model).
  • Token F1: Reached 0.305 (compared to 0.225 for the base model).
  • Valid Envelope: Maintained 100.0% valid command envelope generation.

Further testing on a Q8_0 llama.cpp CPU setup demonstrated a Token F1 of 0.268 and a utility match of 36.0% with efficient decoding speeds (e.g., 32.5 tok/s on 2 threads).

Use Cases

This model is particularly well-suited for:

  • Automated Scripting: Generating specific Bash commands for various tasks.
  • Shell Automation: Integrating into tools that require programmatic shell interaction.
  • Developer Tools: Assisting developers with command-line operations and scripting.

Finetunes
3 models