codelion/Llama-3.2-1B-Instruct-tool-calling-lora

codelion/Llama-3.2-1B-Instruct-tool-calling-lora

🛠️ Tool Calling LoRA with Magpie

This LoRA adapter enhances meta-llama/Llama-3.2-1B-Instruct with tool calling capabilities for code exploration and manipulation. Trained using a hybrid Magpie + real execution approach on diverse coding scenarios.

🎯 Key Features

• —Tool Calling: Teaches models to use development tools effectively
• —Code Exploration: Navigate and understand unfamiliar codebases
• —Real Execution: Training data generated from actual tool execution
• —OpenAI Format: Compatible with OpenAI function calling format
• —Multi-Tool Sequences: Learns to chain multiple tools for complex tasks

📊 Performance Metrics

• —Base Model: meta-llama/Llama-3.2-1B-Instruct
• —Training Method: Standard LoRA fine-tuning
• —LoRA Rank: 64
• —LoRA Alpha: 128
• —Training Samples: 1000
• —Sequence Success Rate: 80.0%
• —Tool Call Accuracy: 80.0%
• —Average Tools per Sequence: 4.0

🔧 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-1B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B-Instruct")

# Load tool calling LoRA adapter
model = PeftModel.from_pretrained(model, "codelion/Llama-3.2-1B-Instruct-tool-calling-lora")

# Example: Use with tool calling prompt
prompt = '''You have access to the following tools:
- list_directory: List contents of a directory
- search_files: Search for files containing specific content
- read_file: Read a single file's contents
- get_file_info: Get file metadata

User: Help me understand how user authentication works in this Flask application
Response:

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

📈 Expected Output Format

The model will generate tool calling sequences in OpenAI format:

{
  "tool_calls": [
    {
      "id": "call_1",
      "type": "function",
      "function": {
        "name": "search_files",
        "arguments": "{\"query\": \"auth\", \"file_types\": [\".py\"]}"
      }
    },
    {
      "id": "call_2",
      "type": "function",
      "function": {
        "name": "read_file",
        "arguments": "{\"path\": \"app.py\"}"
      }
    }
  ]
}

🧪 Training Details

• —Method: Standard LoRA fine-tuning with tool calling data
• —Data Generation: Magpie scenarios + real tool execution
• —Tool Execution: Safe sandbox environment for code exploration
• —Scenario Types: Code exploration, bug hunting, feature addition, refactoring
• —Quality Validation: Minimum tool usage and success rate thresholds

📚 Available Tools

The model is trained to use these development tools:

1. 1.list_directory: Browse project structure
2. 2.Parameters: path (directory to list)

1. 1.search_files: Find files containing specific content
2. 2.Parameters: query, path, file_types, regex

1. 1.read_file: Read complete file contents
2. 2.Parameters: path (file to read)

1. 1.read_multiple_files: Read multiple files at once
2. 2.Parameters: paths (list of files)

1. 1.get_file_info: Get file metadata
2. 2.Parameters: path (file or directory)

1. 1.create_file: Create new files (if safety mode disabled)
2. 2.Parameters: path, content

1. 1.edit_file: Modify existing files (if safety mode disabled)
2. 2.Parameters: path, changes

🎭 Tool Usage Patterns Learned

• —Exploration First: Start with list_directory to understand structure
• —Search Before Read: Use search_files to find relevant files
• —Batch Operations: Use read_multiple_files for related files
• —Progressive Refinement: Start broad, then focus on specific files

🔬 Evaluation

The adapter was evaluated on diverse coding scenarios:

• —Sequence success rate: 80.0%
• —Tool call accuracy: 80.0%
• —Average tools per sequence: 4.0
• —Successful executions: 4/5

🏷️ Tool Usage Distribution

Most frequently used tools during evaluation:

• —get_file_info: 5 uses
• —list_directory: 5 uses
• —read_multiple_files: 5 uses
• —search_files: 5 uses

🏷️ Related

• —Dataset: codelion/Llama-3.2-1B-Instruct-magpie-tool-calling
• —Base Model: meta-llama/Llama-3.2-1B-Instruct
• —Framework: PEFT
• —Training Approach: Magpie + Real Execution

This adapter is part of the [Ellora project](https://github.com/codelion/ellora) - standardized recipes for enhancing LLM capabilities.