naksyu/gemma4-e4b-agent-lime-v1-fft100
The naksyu/gemma4-e4b-agent-lime-v1-fft100 is an experimental 7.9 billion parameter Gemma4 E4B full fine-tune, developed by naksyu, specifically designed for coding-agent behaviors. It excels at structured reasoning, tool-call style formatting, repository/debug workflows, and failure recovery. This model also incorporates a concise Korean-first response tone, making it suitable for agentic tasks requiring precise output and multilingual interaction.
Loading preview...
Overview
naksyu/gemma4-e4b-agent-lime-v1-fft100 is an experimental 7.9 billion parameter full fine-tune of the Gemma4 E4B architecture. This model is a research artifact, focusing on documenting and testing a specific data mixture for agentic capabilities rather than claiming production readiness. It was trained using a full fine-tune method, not LoRA, and primarily covers English and Korean languages.
Key Capabilities
- Coding-Agent Behavior: Designed for practical agent output, including reading tasks carefully, making small verifiable coding/debugging steps, and recovering from errors.
- Tool-Call Formatting: Generates tool-call shaped JSON outputs, intended for use with an external controller for execution.
- Structured Reasoning: Emphasizes structured, visible reasoning to aid in complex task resolution.
- Repository/Debug Workflows: Optimized for tasks related to repository management and debugging, including generating patch summaries.
- Multilingual Interaction: Incorporates a concise Korean-first assistant voice for responses when Korean is used in the prompt.
Good For
- Coding-Agent Prompts: Ideal for scenarios requiring an AI agent to assist with coding tasks, debugging, and structured problem-solving.
- Tool-Use Applications: Suitable for applications where the model needs to generate tool-call shaped outputs for external execution.
- Research and Experimentation: Valuable for researchers exploring agentic behaviors, data mixtures, and fine-tuning techniques for large language models.
Limitations
This model is an experimental snapshot and has not been released with a full benchmark report. Tool-call JSON may occasionally be malformed, and coding/debug suggestions require human validation. The Korean style data is intentionally limited, which may lead to inconsistent tone. It is tuned as text-only, despite the base architecture's potential for multimodal tokens, and safety behavior was not its primary focus.