cagataydev/qwen3-strands-adapter

Qwen3-1.7B Strands Adapter

LoRA adapter that teaches Qwen3-1.7B-4bit comprehensive knowledge of the Strands Agents SDK - without breaking tool calling capabilities.

🎯 What This Does

Transforms base Qwen3-1.7B into a Strands SDK expert that can:

• ✅ Explain Strands architecture and patterns
• ✅ Generate correct installation commands
• ✅ Show working code examples for agent creation
• ✅ Describe tool creation patterns (@tool decorator, hot reload)
• ✅ Explain all model providers (Bedrock, Anthropic, OpenAI, Ollama, MLX)
• ✅ Guide through MLX training pipelines
• ✅ Describe multi-agent systems (Swarm, Workflow)
• ✅ Preserve 100% tool calling capability

📊 Training Results

Metric	Base Model	Trained Model
Strands Knowledge	0/10 (hallucinates)	10/10 (accurate)
Tool Calling	Works	✅ 100% preserved
Training Time	-	14 minutes
Adapter Size	-	19MB

🚀 Quick Start

pip install strands-agents strands-mlx

from strands import Agent
from strands_mlx import MLXModel

# Load base model + this adapter
model = MLXModel(
    model_id="mlx-community/Qwen3-1.7B-4bit",
    adapter_path="cagataycali/qwen3-strands-adapter"  # Your HF repo
)

agent = Agent(model=model)

# Now it knows Strands!
agent("How do I create a Strands tool?")
# → Accurate answer with working code examples

agent("What model providers does Strands support?")
# → Complete list: Bedrock, Anthropic, OpenAI, Ollama, MLX, etc.

📚 Training Details

Dataset:

• 105 question-answer pairs about Strands SDK
• Generated using embedded Strands documentation
• Avg 1,621 tokens per example
• Complete tool call cycles captured (user → assistant(tool_calls) → tool(results) → assistant(final))

Training Configuration:

• Base Model: mlx-community/Qwen3-1.7B-4bit
• Method: LoRA (Low-Rank Adaptation)
• Iterations: 200 (conservative to prevent overfitting)
• Learning Rate: 1e-5
• LoRA Rank: 8
• LoRA Scale: 16.0
• Training Time: 14 minutes on Apple Silicon
• Hardware: M-series Mac

Key Training Strategy:

• Conservative iteration count to avoid catastrophic forgetting
• mask_prompt=True - Only train on assistant completions
• Gradient checkpointing for memory efficiency
• ChatML format with proper tool call structure

🔧 Advanced Usage

With Tools

from strands import Agent
from strands_mlx import MLXModel

model = MLXModel(
    model_id="mlx-community/Qwen3-1.7B-4bit",
    adapter_path="cagataycali/qwen3-strands-adapter"
)

# Tool calling still works perfectly!
agent = Agent(
    model=model,
    tools=["calculator", "shell", "file_read"]
)

agent("Calculate 15 * 7 and explain Strands")
# → Uses calculator AND provides Strands knowledge

Swap Adapters

# Load different adapters for different domains
strands_model = MLXModel(
    "mlx-community/Qwen3-1.7B-4bit",
    adapter_path="cagataycali/qwen3-strands-adapter"
)

coding_model = MLXModel(
    "mlx-community/Qwen3-1.7B-4bit",
    adapter_path="your-coding-adapter"
)

📖 What It Learned

The adapter was trained on comprehensive Strands documentation covering:

1. Installation & Setup

   • pip installation methods
   • Virtual environment setup
   • Model provider configuration

2. Core Concepts

   • Agent creation patterns
   • Tool creation (@tool decorator, hot reload)
   • Session managers
   • System prompts

3. Model Providers

   • Bedrock (AWS)
   • Anthropic
   • OpenAI
   • Ollama (local)
   • MLX (Apple Silicon)
   • Gemini, LiteLLM, GitHub Models

4. Advanced Features

   • Multi-agent systems (Swarm, Workflow)
   • MLX training pipeline
   • MCP server integration
   • Direct tool calls

🎓 Training Your Own Domain Expert

Want to create your own domain-specific adapter? Here's how:

from strands import Agent
from strands_mlx import MLXModel, MLXSessionManager
from strands_mlx.tools import mlx_trainer, dataset_splitter

# 1. Collect training data
session = MLXSessionManager(session_id="my_domain")
model = MLXModel("mlx-community/Qwen3-1.7B-4bit")
agent = Agent(model=model, session_manager=session)

# Have conversations (100-200 examples recommended)
agent("Teach me about X")
# Data auto-saved to ~/.strands/mlx_training_data/my_domain.jsonl

# 2. Split dataset
dataset_splitter(
    input_path="~/.strands/mlx_training_data/my_domain.jsonl",
    train_ratio=0.8,
    valid_ratio=0.1,
    test_ratio=0.1
)

# 3. Train
mlx_trainer(
    action="train",
    config={
        "model": "mlx-community/Qwen3-1.7B-4bit",
        "data": "~/.strands/mlx_training_data/my_domain",
        "iters": 200,
        "learning_rate": 1e-5,
        "batch_size": 1,
        "grad_checkpoint": True
    }
)

# 4. Use it!
trained = MLXModel("mlx-community/Qwen3-1.7B-4bit", adapter_path="./adapter")

🔬 Benchmarks

Question: "How do I create a Strands agent with tools?"

Base Model Response:

I'll help you create a strands agent. Let me use the system_monitor tool...
<tool_call>{"name": "system_monitor", ...}</tool_call>  ← Hallucinated tool

Trained Model Response:

from strands import Agent

agent = Agent(
    system_prompt="You are a helpful assistant",
    tools=["calculator", "shell", "file_read"]
)

agent("Calculate 2+2")

📦 Files in This Adapter

• adapters.safetensors - Trained LoRA weights
• adapter_config.json - LoRA configuration
• 0000100_adapters.safetensors - Checkpoint at iter 100
• 0000200_adapters.safetensors - Checkpoint at iter 200

🛠️ Requirements

pip install strands-agents>=0.2.0 strands-mlx>=0.2.1

📝 License

Same as base model (Apache 2.0)

🙏 Acknowledgments

• Built with Strands Agents SDK
• Base model: Qwen3-1.7B-4bit
• Training framework: MLX

🔗 Links

• Strands SDK: https://github.com/strands-agents/sdk-python
• Strands MLX: https://github.com/cagataycali/strands-mlx
• Base Model: https://huggingface.co/mlx-community/Qwen3-1.7B-4bit