LLM-OS-Models/LFM2.5-1.2B-Terminal-SFT-1Epoch-LiquidCLI-TemplateHoldout

TEXT GENERATIONConcurrency Cost:1Model Size:1.2BQuant:BF16Ctx Length:32kPublished:May 7, 2026Architecture:Transformer Cold

LLM-OS-Models/LFM2.5-1.2B-Terminal-SFT-1Epoch-LiquidCLI-TemplateHoldout is a 1.2 billion parameter model based on LiquidAI/LFM2.5-1.2B-Instruct, specifically fine-tuned for terminal automation tasks. It generates subsequent terminal commands in JSON format based on user input and previous terminal states. This model is optimized for cost-effective and fast inference in specific acceleration paths, demonstrating a conservative approach to command generation.

Loading preview...

Overview

This model, LLM-OS-Models/LFM2.5-1.2B-Terminal-SFT-1Epoch-LiquidCLI-TemplateHoldout, is a 1.2 billion parameter model derived from LiquidAI/LFM2.5-1.2B-Instruct. It is specialized for terminal automation, designed to generate the next command in JSON format given a task description and the current terminal state. The model was trained for 1 epoch using Liquid-CLI style preprocessing and training, with a chat-template aligned holdout split.

Key Capabilities & Characteristics

  • Terminal Command Generation: Generates JSON-formatted commands for terminal automation.
  • JSON Output: Recommends a specific JSON output format including analysis, plan, commands (with keystrokes and duration), and task_complete.
  • Conservative Command Generation: Tends to issue correct commands conservatively rather than many incorrect ones.
  • Efficiency: Offers fast inference at 0.085 sec/step, making it suitable for iterative evaluation and RL experiments.
  • Base Model: Built upon the LFM series, which shows significant SFT performance gains relative to its base score.

Performance & Limitations

Evaluated on the corrected TB2-lite replay set, the model achieved a score of 28.10 (Command F1) and a rank of 36 / 56. Key metrics include:

  • Command F1: 0.2810
  • Valid JSON: 50.5%
  • First command exact: 29.4%

Limitations include:

  • Lower Recall: May omit some necessary commands.
  • JSON Format Failures: Requires parsing validation and retries before execution due to a 50.5% valid JSON rate.
  • Specialized Use: This model is specifically for automated terminal operation assistance and does not guarantee general conversation or reasoning performance.
  • Safety: Generated commands require safety measures like sandboxing, allowlisting, or human review before execution.

Use Cases

  • Automated Terminal Operations: Ideal for scenarios requiring automated command execution in a terminal environment.
  • Cost-Effective Inference: Suitable for applications where fast inference and cost efficiency are critical in specific acceleration paths.
  • RL Experimentation: Its efficiency makes it a good candidate for reinforcement learning experiments focused on improving terminal interaction.

Finetunes
1 models