gemma-4-E4B-it-open-doc

A LoRA adapter for MCP tool calling, trained by colab-finetune-mcp through its teacher-driven pipeline: reverse-generate the data, filter it for quality, then run SFT.

• Base model: google/gemma-4-E4B-it
• Method: LoRA (r=16, epochs=3.0, learning_rate=0.0002)
• Tools (from the MCP server): list_documents, read_document, create_document, list_templates, create_from_template, write_document, rename_document, duplicate_document, delete_document, read_text, write_text, add_comment, list_themes, read_theme, list_assets, upload_asset, find_asset_usages, delete_asset, list_folders, create_folder, file_document, check_layout, render_page, export_document, diff_document, import_markdown

Evaluation (held-out test set)

Metric	Base (before)	Fine-tuned (after)
Tool-call rate	97.0%	100.0%
Function-name accuracy	60.0%	99.3%
Exact-match rate	56.3%	91.1%
No-tool accuracy	100.0%	100.0%

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "google/gemma-4-E4B-it"
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "Simon-Liu/gemma-4-E4B-it-open-doc")
tokenizer = AutoTokenizer.from_pretrained("Simon-Liu/gemma-4-E4B-it-open-doc")

The model is trained to emit a single <tool_call>{"name": "<tool>", "arguments": {...}}</tool_call> when the user's request needs a tool, and to answer in plain text when it does not.