Qwen2.5-MCP-Tool-Calling

🛠️ Fine-tuned Tool-Calling LLM (LoRA + Merged) for Geospatial Operations

This repository contains a fine-tuned Large Language Model (LLM) capable of structured tool/function calling, optimized for integration with backend services such as Model Context Protocol (MCP) for geospatial, file analysis, and automation tasks.

Note:This repo only has the lora adapter of the model.You can merge it with the Qwen2.5-Coder-1.5B-Instruct base model.

Model Structure

lora_model/

LoRA (Low-Rank Adaptation) adapters and tokenizer files for efficient fine-tuning and inference.

• adapter_model.safetensors, adapter_config.json — LoRA adapter weights and config.
• tokenizer.json, tokenizer_config.json, special_tokens_map.json, vocab.json — Tokenizer files.
• chat_template.jinja, merges.txt, added_tokens.json — (If applicable) chat formatting and tokenization.

merged_model/

The base model merged with LoRA adapters for direct use without extra configuration.

• model-00001-of-00002.safetensors, model-00002-of-00002.safetensors — Model weights (split).
• config.json, generation_config.json — Model and generation configs.
• tokenizer.json, tokenizer_config.json, special_tokens_map.json, vocab.json — Tokenizer files.
• chat_template.jinja, merges.txt, added_tokens.json — (If applicable) chat formatting and tokenization.

---

How to Use

Loading the Merged Model

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("path/to/merged_model")
tokenizer = AutoTokenizer.from_pretrained("path/to/merged_model")

# Example inference
input_text = "Analyze the file C:/data/image.tif"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0]))

Loading with LoRA Adapter (Optional)

If you want to use the LoRA adapter separately:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("your/base-model")
tokenizer = AutoTokenizer.from_pretrained("path/to/lora_model")
model = PeftModel.from_pretrained(base_model, "path/to/lora_model")

Example inference...

Intended Use Natural language to tool-calling JSON conversion for backend automation.

Geospatial and image operations via MCP or custom tools.

Easily extensible for new tool schemas.

Example Input/Output Prompt:

Crop the image C:/images/sample.tif to bounding box [xmin, ymin, xmax, ymax].

Model Output:

TOOL_NEEDED: crop_image
PARAMS: {"filepath": "C:/images/sample.tif", "minx": ..., "miny": ..., "maxx": ..., "maxy": ...}

Training Details

Fine-tuned for structured function-calling on domain-specific data.

Supports both merged and LoRA-adapter workflows.

For technical backend details, see the project repository.

Citation

@misc{finetuned_llm_mcp_2025, author = {İsmail Emre Candan}, title = {Fine-tuned Tool-Calling LLM (LoRA + Merged)}, year = {2025}, url = {https://github.com/EmreCandan0/fine-tuned-llm-tool-calling} }