README.md

---
library_name: transformers
license: mit
base_model: Qwen/Qwen3-4B-Instruct-2507
tags:
  - code
  - agent
  - tool-calling
  - distillation
  - qwen3
  - gguf
  - llama-cpp
language:
  - en
pipeline_tag: text-generation
---

<div align="center">
  <img src="assets/loco_operator.png" width="55%" alt="LocoOperator" />
</div>

<br>

<div align="center">

[![MODEL](https://img.shields.io/badge/Model-FFB300?style=for-the-badge&logo=huggingface&logoColor=white)](https://huggingface.co/LocoreMind/LocoOperator-4B)
[![Blog](https://img.shields.io/badge/Blog-4285F4?style=for-the-badge&logo=google-chrome&logoColor=white)](https://locoremind.com/blog/loco-operator)
[![GitHub](https://img.shields.io/badge/GitHub-181717?style=for-the-badge&logo=github&logoColor=white)](https://github.com/LocoreMind/LocoOperator)
[![Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&logoColor=white)](https://colab.research.google.com/github/LocoreMind/LocoOperator/blob/main/LocoOperator_4B.ipynb)

</div>

## Introduction

**LocoOperator-4B** is a 4B-parameter tool-calling agent model trained via knowledge distillation from **Qwen3-Coder-Next** inference traces. It specializes in multi-turn codebase exploration — reading files, searching code, and navigating project structures within a Claude Code-style agent loop. Designed as a local sub agent, it runs via llama.cpp at zero API cost.

|  | LocoOperator-4B |
|:--|:--|
| **Base Model** | [Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507) |
| **Teacher Model** | Qwen3-Coder-Next |
| **Training Method** | Full-parameter SFT (distillation) |
| **Training Data** | 170,356 multi-turn conversation samples |
| **Max Sequence Length** | 16,384 tokens |
| **Training Hardware** | 4x NVIDIA H200 141GB SXM5 |
| **Training Time** | ~25 hours |
| **Framework** | MS-SWIFT |

## Key Features

- **Tool-Calling Agent**: Generates structured `<tool_call>` JSON for Read, Grep, Glob, Bash, Write, Edit, and Task (subagent delegation)
- **100% JSON Validity**: Every tool call is valid JSON with all required arguments — outperforming the teacher model (87.6%)
- **Local Deployment**: GGUF quantized, runs on Mac Studio via llama.cpp at zero API cost
- **Lightweight Explorer**: 4B parameters, optimized for fast codebase search and navigation
- **Multi-Turn**: Handles conversation depths of 3–33 messages with consistent tool-calling behavior

## Performance

Evaluated on 65 multi-turn conversation samples from diverse open-source projects (scipy, fastapi, arrow, attrs, gevent, gunicorn, etc.), with labels generated by Qwen3-Coder-Next.

### Core Metrics

| Metric | Score |
|:-------|:-----:|
| **Tool Call Presence Alignment** | **100%** (65/65) |
| **First Tool Type Match** | **65.6%** (40/61) |
| **JSON Validity** | **100%** (76/76) |
| **Argument Syntax Correctness** | **100%** (76/76) |

The model perfectly learned *when* to use tools vs. when to respond with text (100% presence alignment). Tool type mismatches are between semantically similar tools (e.g. Grep vs Read) — different but often valid strategies.

### Tool Distribution Comparison

<div align="center">
  <img src="assets/tool_distribution.png" width="80%" alt="Tool Distribution Comparison" />
</div>

### JSON & Argument Syntax Correctness

| Model | JSON Valid | Argument Syntax Valid |
|:------|:---------:|:--------------------:|
| **LocoOperator-4B** | 76/76 (100%) | 76/76 (100%) |
| Qwen3-Coder-Next (teacher) | 89/89 (100%) | 78/89 (87.6%) |

> LocoOperator-4B achieves perfect structured output. The teacher model has 11 tool calls with missing required arguments (empty `arguments: {}`).

## Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "LocoreMind/LocoOperator-4B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the messages
messages = [
    {
        "role": "system",
        "content": "You are a read-only codebase search specialist.\n\nCRITICAL CONSTRAINTS:\n1. STRICTLY READ-ONLY: You cannot create, edit, delete, move files, or run any state-changing commands. Use tools/bash ONLY for reading (e.g., ls, find, cat, grep).\n2. EFFICIENCY: Spawn multiple parallel tool calls for faster searching.\n3. OUTPUT RULES: \n   - ALWAYS use absolute file paths.\n   - STRICTLY NO EMOJIS in your response.\n   - Output your final report directly. Do not use colons before tool calls.\n\nENV: Working directory is /Users/developer/workspace/code-analyzer (macOS, zsh)."
    },
    {
        "role": "user",
        "content": "Analyze the Black codebase at `/Users/developer/workspace/code-analyzer/projects/black`.\nFind and explain:\n1. How Black discovers config files.\n2. The exact search order for config files.\n3. Supported config file formats.\n4. Where this configuration discovery logic lives in the codebase.\n\nReturn a comprehensive answer with relevant code snippets and absolute file paths."
    }
]

# prepare the model input
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512,
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

content = tokenizer.decode(output_ids, skip_special_tokens=True)
print(content)
```

## Local Deployment

For GGUF quantized deployment with llama.cpp, hybrid proxy routing, and batch analysis pipelines, refer to our [GitHub repository](https://github.com/LocoreMind/LocoOperator).

## Training Details

| Parameter | Value |
|:----------|:------|
| Base model | Qwen3-4B-Instruct-2507 |
| Teacher model | Qwen3-Coder-Next |
| Method | Full-parameter SFT |
| Training data | 170,356 samples |
| Hardware | 4x NVIDIA H200 141GB SXM5 |
| Parallelism | DDP (no DeepSpeed) |
| Precision | BF16 |
| Epochs | 1 |
| Batch size | 2/GPU, gradient accumulation 4 (effective batch 32) |
| Learning rate | 2e-5, warmup ratio 0.03 |
| Max sequence length | 16,384 tokens |
| Template | qwen3_nothinking |
| Framework | MS-SWIFT |
| Training time | ~25 hours |
| Checkpoint | Step 2524 |

## Known Limitations

- First-tool-type match is 65.6% — the model sometimes picks a different (but not necessarily wrong) tool than the teacher
- Tends to under-generate parallel tool calls compared to the teacher (76 vs 89 total calls across 65 samples)
- Preference for Bash over Read may indicate the model defaults to shell commands where file reads would be more appropriate
- Evaluated on 65 samples only; larger-scale evaluation needed

## License

MIT

## Acknowledgments

- [Qwen Team](https://huggingface.co/Qwen) for the Qwen3-4B-Instruct-2507 base model
- [MS-SWIFT](https://github.com/modelscope/ms-swift) for the training framework
- [llama.cpp](https://github.com/ggerganov/llama.cpp) for efficient local inference
- [Anthropic](https://www.anthropic.com/) for the Claude Code agent loop design that inspired this work