Compactbot/liquidagent-1.2b

LiquidAgent-1.2B Overview

LiquidAgent-1.2B is a 1.2 billion parameter language model developed by Compactbot, fine-tuned from LiquidAI/LFM2.5-1.2B-Thinking. This model is specifically adapted for agentic tool-use, enabling it to generate tool calls, interpret their results, and integrate this information into subsequent reasoning steps. It leverages a hybrid architecture combining LIV-conv and GQA blocks, supporting a 32,768 token context length.

Key Capabilities

• Agentic Tool-Use: Learns to emit tool calls and process tool results, facilitating interaction with external systems.
• Reasoning with Tool Outputs: Integrates tool feedback into its reasoning process for more effective problem-solving.
• Efficient Architecture: Built on a 1.17B-parameter hybrid architecture (10 LIV-conv + 6 GQA blocks) with built-in thinking capabilities.
• High Context Length: Supports a 32,768 token context window, allowing for processing of extensive interactions.

Good For

• Automated Task Execution: Ideal for applications requiring models to interact with tools to complete tasks.
• Complex Problem Solving: Suitable for scenarios where external information retrieval or action is necessary.
• Agent Development: A strong base for building and experimenting with AI agents that require tool integration.

This model addresses a critical oversight in previous agentic fine-tunes by correctly parsing and utilizing toolResult messages, ensuring comprehensive training on agent traces. While trained on a relatively small dataset of 379 examples, it demonstrates coherent and on-topic generation for agentic workflows.