LocoTrainer-4B is a 4B-parameter MS-SWIFT domain expert agent distilled from Qwen3-Coder-Next. Designed to analyze MS-SWIFT codebases and generate comprehensive markdown reports — combining tool-calling capabilities with deep framework knowledge.

🎬 Demo

LocoTrainer analyzing MS-SWIFT codebase with LocoTrainer-4B model via vLLM

📋 Table of Contents

• 📰 News & Updates
• 📝 Introduction
• ✨ Key Features
• 🏗️ Architecture
• 📈 Performance
• 🚀 Quick Start
• 🔧 Usage Examples
• 📁 Project Structure
• ⚠️ Known Limitations
• 📄 License
• 🙏 Acknowledgments

📰 News & Updates

• [2026-03-13] 🚀 LocoTrainer-4B and LocoTrainer framework released.

📝 Introduction

LocoTrainer-4B is a specialized code analysis agent trained via knowledge distillation from Qwen3-Coder-Next. Unlike general-purpose code agents, it combines multi-turn tool-calling capabilities with deep MS-SWIFT framework knowledge, enabling it to generate comprehensive codebase analysis reports without requiring a separate reasoning model.

	LocoTrainer-4B
Base Model	Qwen3-4B-Instruct-2507
Teacher Model	Qwen3-Coder-Next
Training Method	Full-parameter SFT (distillation)
Training Data	361,830 samples (agent trajectory + MS-SWIFT knowledge + project paths)
Max Sequence Length	32,768 tokens
Training Hardware	8x NVIDIA H100 80GB
Training Time	~25 hours
Framework	MS-SWIFT

✨ Key Features

• 🎯 MS-SWIFT Domain Expert: Trained on MS-SWIFT documentation, CLI parameters, and project structure — answers framework questions accurately without hallucination
• 🔧 Tool-Calling Agent: Generates structured <tool_call> JSON for Read, Grep, Glob, Bash, and Write tools
• 📊 End-to-End Reports: From a single question to a complete, well-structured markdown analysis report
• 🏠 Local Deployment: GGUF quantized, runs on Mac Studio via llama.cpp at zero API cost
• 📏 Long Context: 32K training covers 90% of long-context analysis scenarios
• 🔄 Auto Clone: Automatically clones ms-swift on first run — no manual setup needed

🏗️ Architecture

LocoTrainer consists of two components: the agent framework (this repo) and the LocoTrainer-4B model.

User Query
    │
    ▼
LocoTrainer Framework
    ├── build_user_query()     # injects absolute paths
    ├── get_system_reminder()  # simulates Claude Code environment
    └── Agent Loop
            │
            ▼
    LocoTrainer-4B (or any OpenAI-compatible model)
            │
            ├── <tool_call> Read / Grep / Glob / Bash
            │       │
            │       ▼
            │   Real Filesystem (ms-swift codebase)
            │       │
            │       ▼
            └── <tool_response> → next turn
                    │
                    ▼
            output/output.md   (final markdown report)
            output/trajectory.json  (full conversation log)

The framework simulates a Claude Code-style agent environment, which is exactly what LocoTrainer-4B was trained on — ensuring maximum compatibility between the model and the runtime.

📈 Performance

Evaluated on MS-SWIFT codebase analysis tasks across 3 test iterations.

Agent Loop Reliability

Metric	Test 1 (Relative Paths)	Test 2 (Absolute Paths)	Test 3 (Absolute + Tolerant Args)
Read success rate	0%	100%	100%
Write success rate	—	0%	100%
Turns used	15 (limit)	15 (limit)	9
Tool calls	34	27	13
output.md generated	✗	✗	✓ (225 lines)

Key Design Insight

Absolute paths in user content + tolerant tool argument parsing = reliable agent behavior.

This mirrors Claude Code's own design: the system always provides full absolute paths so the model never has to guess.

🚀 Quick Start

Option 1: One-Line Setup (Recommended)

We provide automated setup scripts for different scenarios:

For most users (cloud API):

curl -O https://raw.githubusercontent.com/LocoreMind/LocoTrainer/main/scripts/setup_locotrainer.sh
chmod +x setup_locotrainer.sh
./setup_locotrainer.sh

For GPU users (local vLLM + LocoTrainer-4B):

curl -O https://raw.githubusercontent.com/LocoreMind/LocoTrainer/main/scripts/setup_locotrainer_vllm.sh
chmod +x setup_locotrainer_vllm.sh
./setup_locotrainer_vllm.sh

For developers (from source):

git clone https://github.com/LocoreMind/LocoTrainer.git
cd LocoTrainer
./scripts/setup_locotrainer_dev.sh

See scripts/README.md for detailed documentation.

---

Option 2: Manual Installation

Prerequisites

• An OpenAI-compatible API key (DashScope, OpenRouter, or local llama.cpp)

Install

pip install locotrainer

Configure

export LOCOTRAINER_API_KEY=your-api-key
export LOCOTRAINER_BASE_URL=https://api.openai.com/v1  # or your endpoint
export LOCOTRAINER_MODEL=gpt-4o

Or use a .env file (clone the repo for the template):

git clone https://github.com/LocoreMind/LocoTrainer.git
cd LocoTrainer && cp .env.example .env
# Edit .env with your API key

Run (auto-clones ms-swift on first use)

locotrainer run -q "What are the default LoRA settings in ms-swift?"
# → output/output.md

With LocoTrainer-4B via llama.cpp

# Start local server
./llama-server -m LocoTrainer-4B.gguf --ctx-size 51200 --port 8080

# Configure and run
export LOCOTRAINER_BASE_URL=http://localhost:8080/v1
export LOCOTRAINER_MODEL=LocoTrainer-4B
export LOCOTRAINER_API_KEY=local

locotrainer run -q "How does ms-swift implement GRPO training?"

With DashScope (Qwen3-Coder-Next)

LOCOTRAINER_API_KEY=sk-your-dashscope-key
LOCOTRAINER_BASE_URL=https://dashscope-intl.aliyuncs.com/compatible-mode/v1
LOCOTRAINER_MODEL=qwen3-coder-next
LOCOTRAINER_ENABLE_THINKING=true

🔧 Usage Examples

Analyze a specific topic

locotrainer run -q "What are all supported training methods in ms-swift and their differences?"

Analyze a different codebase

locotrainer run \
  -q "How does the authentication system work?" \
  -c /path/to/other/project \
  -o ./reports

Full CLI options

locotrainer run \
  -q "Your question here" \
  -c /path/to/codebase \    # optional, default: auto-clone ms-swift
  -o ./output \             # optional, default: ./output
  -m model-name \           # optional, overrides .env
  --max-turns 20 \          # optional, default: 20
  --quiet                   # suppress verbose output

📁 Project Structure

LocoTrainer/
├── pyproject.toml              # uv project config, pip installable
├── .env.example                # Configuration template
├── .env                        # User config (gitignored)
└── src/locotrainer/
    ├── __init__.py             # Package version
    ├── prompts.py              # SYSTEM_PROMPT + get_system_reminder()
    ├── tools.py                # ToolExecutor (Read/Grep/Glob/Write/Bash)
    ├── agent.py                # Agent loop
    ├── config.py               # Config dataclass, .env loading
    ├── repo.py                 # Auto-clone ms-swift logic
    └── cli.py                  # Click CLI: `locotrainer run`

⚙️ Configuration

Env Variable	Default	Description
LOCOTRAINER_API_KEY	(required)	API key (falls back to OPENAI_API_KEY)
LOCOTRAINER_BASE_URL	https://api.openai.com/v1	OpenAI-compatible endpoint
LOCOTRAINER_MODEL	gpt-4o	Model name
LOCOTRAINER_MAX_TURNS	20	Max agent loop turns
LOCOTRAINER_MAX_TOKENS	8192	Max tokens per response
LOCOTRAINER_ENABLE_THINKING	false	Enable thinking mode (Qwen3)
LOCOTRAINER_CODEBASE	.	Codebase path (triggers auto-clone if default)
LOCOTRAINER_OUTPUT_DIR	./output	Output directory

Training Details

📋 Click to expand full training configuration

Parameter	Value
Base model	Qwen3-4B-Instruct-2507
Teacher model	Qwen3-Coder-Next
Method	Full-parameter SFT
Training data	361,830 samples
Data composition	Agent trajectory + MS-SWIFT knowledge + project structure paths
Hardware	8x NVIDIA H100 80GB
DeepSpeed	ZeRO-2
Precision	BF16
Epochs	1
Max sequence length	32,768 tokens
Attention	Flash Attention 2
Kernel optimization	Liger Kernel
Learning rate	1e-5, warmup ratio 0.05
Batch size	1/GPU, gradient accumulation 4 (effective batch 32)
Template	qwen3_nothinking
Framework	MS-SWIFT
Training time	~25 hours

⚠️ Known Limitations

• Specialized for MS-SWIFT; performance on unrelated codebases is untested
• 4B parameters — complex multi-hop reasoning may require a larger model
• MS-SWIFT project structure knowledge reflects the training data snapshot; may drift as the framework evolves

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• 🤖 Qwen Team for the Qwen3-4B-Instruct-2507 base model
• 🛠️ MS-SWIFT for the training framework and the codebase this model specializes in
• 🦙 llama.cpp for efficient local inference
• 🤖 Anthropic for the Claude Code agent loop design that inspired this work