Turbocharge LLaMA Fine-Tuning with Tuna-Asyncio: A No-Code Solution

# Clone the Tuna-Asyncio with LLaMA repository
git clone https://gitlab.com/krafi/tuna-asyncio-with-llama.git
cd tuna-asyncio-with-llama

# Install required Python packages
pip install -r requirements.txt

Common dependencies include:

• pandas - For CSV processing
• asyncio - For async operations (built-in)
• Your LLaMA client library

Ensure you have a local LLaMA instance running. You can use:

• LLM.cpp for local inference
• Ollama for easy local setup
• Any LLaMA-compatible API endpoint

Create a file named chunk.csv in the project directory. Each row should contain the text data you want to generate Q&A pairs from.

Example chunk.csv format:

chunk
"How to reset your password: Go to settings, click on security, select reset password, enter your current password, then create a new one."
"The weather today is sunny with a temperature of 72°F. It's a perfect day for outdoor activities."
"Python list comprehension is a concise way to create lists. Example: [x for x in range(10) if x % 2 == 0]"

Open main.py and configure your LLaMA endpoint:

# Example configuration in main.py
LLAMA_ENDPOINT = "http://localhost:8080/v1/chat/completions"
MODEL_NAME = "llama-3-8b"  # Or your model name

Execute the main script to generate your dataset:

python main.py

You’ll see progress as each chunk is processed:

Processing chunk 1/100...
Processing chunk 2/100...
...
✅ Done! Generated 500 Q&A pairs
📁 Output saved to: output_alpaca.json

Your output_alpaca.json will look like this:

[
  {
    "instruction": "How do I reset my password?",
    "input": "",
    "output": "To reset your password: 1. Go to settings 2. Click on security 3. Select reset password 4. Enter your current password 5. Create a new password"
  },
  {
    "instruction": "What is the weather like today?",
    "input": "",
    "output": "The weather today is sunny with a temperature of 72°F. It's perfect for outdoor activities."
  }
]

1. Open the Google Colab notebook
2. Upload your output_alpaca.json to the LLaMA-Factory/data directory
3. Update identity.json to include your dataset:

{
  "identity": {
    "file_name": "identity.json"
  },
  "alpaca_en_demo": {
    "file_name": "alpaca_en_demo.json"
  },
  "output_alpaca": {
    "file_name": "output_alpaca.json"
  },
  "alpaca_zh_demo": {
    "file_name": "alpaca_zh_demo.json"
  }
}

4. Run the notebook cells to start fine-tuning

For local fine-tuning, use LLaMA-Factory:

# Clone LLaMA-Factory
git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory

# Copy your dataset
cp /path/to/output_alpaca.json data/

# Start fine-tuning
python src/train.py \
    --model_name_or_path llama-3-8b \
    --dataset output_alpaca \
    --output_dir ./trained_model \
    --num_train_epochs 3 \
    --per_device_train_batch_size 4

Error: “Cannot connect to LLaMA instance”

Solution:

• Verify your LLaMA instance is running
• Check the endpoint URL in main.py
• Ensure no firewall is blocking the connection

# Test your endpoint
curl http://localhost:8080/health

Generated Q&A pairs are not relevant

Solution:

• Provide more detailed input text in your CSV
• Try a larger or more capable LLaMA model
• Add more examples to improve context

Out of memory during generation

Solution:

• Process CSV in smaller batches
• Reduce concurrent requests
• Use a smaller LLaMA model for generation

Output JSON is malformed

Solution:

• Check your CSV for special characters
• Ensure proper escaping in input text
• Validate JSON output with a linter