# 🤖 MCP-Agent-1.7B

> The first open-source small language model fine-tuned to natively speak the **Model Context Protocol (MCP)**.

## What Is This?

MCP-Agent-1.7B is [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) fine-tuned with LoRA to:

- 🔧 **Call tools** using MCP protocol (JSON-RPC format)
- 🔗 **Plan multi-step tool chains** with DAG dependencies
- ❓ **Ask clarifying questions** when info is missing (instead of hallucinating)
- 🛡️ **Refuse dangerous requests** (e.g., "delete all files")

## Training

| Detail | Value |
|--------|-------|
| Base model | [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) (2B params) |
| Method | LoRA SFT (rank=16, all-linear) |
| Dataset | [muhammadtlha944/mcp-agent-training-data](https://huggingface.co/datasets/muhammadtlha944/mcp-agent-training-data) (16.5K examples) |
| Epochs | 3 |
| Learning rate | 2e-4 (cosine schedule) |
| Effective batch size | 16 |
| Precision | fp16 |
| Cost | **$0** (trained on free Kaggle/Colab GPU) |

## 🚀 Train It Yourself (FREE)

### Option 1: Kaggle (Recommended — 30 free GPU hrs/week)
1. Go to [kaggle.com/code](https://kaggle.com/code) → **New Notebook**
2. **Settings** → Accelerator → **GPU T4×2** → Internet → **ON**
3. Copy-paste the code from [`MCP_Agent_1_7B_Kaggle.ipynb`](./MCP_Agent_1_7B_Kaggle.ipynb)
4. Run all cells → wait ~2 hours → model auto-pushes to Hub

### Option 2: Google Colab (Backup — free T4)
1. Open [colab.research.google.com](https://colab.research.google.com)
2. **Runtime** → Change runtime type → **T4 GPU**
3. Copy-paste from [`MCP_Agent_1_7B_Training.ipynb`](./MCP_Agent_1_7B_Training.ipynb)
4. Run all cells → wait ~2 hours

## Quick Start

```python
from transformers import pipeline

pipe = pipeline("text-generation", model="muhammadtlha944/MCP-Agent-1.7B")

messages = [
    {"role": "system", "content": "You are an MCP agent with tools: github_search, read_file, shell_exec."},
    {"role": "user", "content": "Find all Python files that import pandas"}
]

response = pipe(messages, max_new_tokens=512)
print(response[0]['generated_text'][-1]['content'])
```

## Research Background

This project is informed by:
- **STAR Framework** ([arxiv 2602.03022](https://arxiv.org/abs/2602.03022)) — proved Qwen3-1.7B beats Llama-3.1-8B at function calling
- **TinyAgent** ([arxiv 2409.00608](https://arxiv.org/abs/2409.00608)) — proved 1.1B models can match GPT-4 at focused tool calling
- **LoRA Without Regret** — all-linear LoRA matches full fine-tuning quality

## License

Apache 2.0 (same as base model)

## Author

Built by [Muhammad Talha](https://huggingface.co/muhammadtlha944) — learning ML by building real projects.