Qwen2.5-Coder-32B-Glaive-ToolCall

Model Description

This model is a fine-tuned version of Qwen/Qwen2.5-Coder-32B-Instruct specifically enhanced for tool calling capabilities. The model has been trained using the Glaive Function Calling v2 dataset (glaiveai/glaive-function-calling-v2) to significantly improve its ability to understand, generate, and execute function calls in various programming and automation contexts.

Model Details

• Base Model: Qwen/Qwen2.5-Coder-32B-Instruct
• Model Type: Large Language Model (LLM) with enhanced tool calling capabilities
• Architecture: Transformer-based decoder model
• Parameters: 32 billion parameters
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Training Dataset: glaive-function-calling-v2
• Language Support: Multilingual

Training Configuration

• Fine-tuning Type: LoRA with rank 8, alpha 16
• Training Epochs: 3.0
• Learning Rate: 5e-5 with cosine scheduler
• Batch Size: 2 per device with 8 gradient accumulation steps
• Context Length: 2048 tokens
• Optimizer: AdamW
• Precision: BF16
• Max Samples: 100,000

Enhanced Capabilities

Tool Calling Improvements

This model demonstrates significant improvements in:

1. Function Schema Understanding: Enhanced ability to parse and understand complex function signatures and parameter requirements
2. Context-Aware Tool Selection: Improved decision-making for selecting appropriate tools based on user queries
3. Parameter Extraction: Better extraction and formatting of function parameters from natural language inputs
4. Multi-step Tool Orchestration: Enhanced capability to chain multiple tool calls for complex tasks
5. Error Handling: Improved error detection and recovery in tool calling scenarios

Key Features

• Robust JSON Generation: Produces well-formatted JSON for function calls with proper schema adherence
• Natural Language Integration: Seamlessly integrates tool calls within conversational responses
• Code Generation with Tools: Enhanced ability to generate code that incorporates external tool usage
• API Integration: Improved understanding of REST APIs, GraphQL, and other web service interfaces

Use Cases

This model is particularly well-suited for:

• AI Assistants: Building conversational AI that can interact with external systems
• Automation Workflows: Creating intelligent automation scripts with dynamic tool usage
• Code Generation: Generating code that integrates with APIs and external services
• Data Processing: Automating data analysis and processing tasks with appropriate tools
• System Integration: Building bridges between different software systems and services

Usage Example

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load the model and tokenizer
model_name = "RekklesAI/Qwen2.5-Coder-32B-Glaive-ToolCall"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Example prompt for tool calling
prompt = """You have access to a weather API. Help me get the current weather for New York City.

Available tools:
- get_weather(location: str, units: str = "metric") -> dict

User: What's the weather like in New York City?"""

# Generate response
inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
    outputs = model.generate(
        inputs.input_ids,
        max_new_tokens=512,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Performance Metrics

The model shows significant improvements in tool calling benchmarks:

• Function Call Accuracy: Enhanced precision in generating syntactically correct function calls
• Parameter Extraction: Improved accuracy in extracting relevant parameters from user queries
• Tool Selection: Better performance in selecting appropriate tools for given tasks
• JSON Formatting: Reduced errors in JSON structure and formatting

Training Loss

The following chart shows the training loss progression during the fine-tuning process:

Training loss curve demonstrating stable convergence over 3 epochs with the Glaive Function Calling v2 dataset.

Limitations

• The model's tool calling capabilities are primarily trained on the patterns present in the Glaive Function Calling v2 dataset
• Performance may vary for highly specialized or domain-specific tools not represented in the training data
• Like all LLMs, the model may occasionally generate plausible-sounding but incorrect tool calls
• The model requires careful prompt engineering for optimal tool calling performance

Ethical Considerations

• Tool Safety: Users should implement proper validation and sandboxing when allowing the model to execute actual tool calls
• Access Control: Implement appropriate access controls and permissions for tools accessible to the model
• Data Privacy: Be mindful of sensitive data that might be passed through tool calls
• Monitoring: Implement logging and monitoring for tool usage in production environments

Training Data

The model was fine-tuned using the Glaive Function Calling v2 dataset (glaiveai/glaive-function-calling-v2), a comprehensive and high-quality dataset specifically designed for training language models in function calling capabilities.

Dataset Overview

• Dataset Size: 113,000 training examples
• Format: JSON with structured conversations
• Language: English
• License: Apache 2.0
• Source: Glaive AI

Dataset Characteristics

The Glaive Function Calling v2 dataset is meticulously curated to provide diverse and realistic function calling scenarios:

Conversation Structure

• System Messages: Define the assistant's role and available functions with detailed schemas
• Multi-turn Dialogues: Natural conversations between users and AI assistants
• Function Calls: Properly formatted JSON function invocations
• Function Responses: Realistic API responses and result handling
• Error Scenarios: Examples of graceful error handling and capability limitations

Function Diversity

The dataset covers a wide range of function types and use cases:

• Utility Functions: Email sending, calendar management, password generation
• Data Retrieval: News headlines, stock prices, weather information
• Computational Tasks: Mathematical calculations, unit conversions, data analysis
• Search Operations: Movie searches, book lookups, general information retrieval
• Communication Tools: Contact management, messaging systems
• Financial Services: Exchange rates, loan calculations, investment data
• Content Creation: Text generation, formatting, summarization

Quality Features

1. Realistic Scenarios: Conversations mirror real-world user interactions with AI assistants
2. Proper Error Handling: Examples of polite refusals when functions are unavailable
3. Parameter Validation: Correct handling of required and optional function parameters
4. Context Awareness: Functions are called appropriately based on conversation context
5. Natural Language Integration: Seamless integration of function results into conversational responses

Training Examples Include:

• Single Function Calls: Simple, direct function invocations
• Multi-step Workflows: Complex scenarios requiring multiple function calls
• Parameter Extraction: Converting natural language requests into structured function parameters
• Response Formatting: Presenting function results in user-friendly formats
• Capability Boundaries: Clear communication of system limitations

Dataset Impact on Model Performance

This carefully curated dataset enables the model to:

• Understand Function Schemas: Parse and comprehend complex function definitions
• Extract Parameters: Accurately identify and format required function arguments from user queries
• Generate Valid JSON: Produce syntactically correct function calls
• Handle Edge Cases: Manage scenarios where requested functions are unavailable
• Maintain Conversational Flow: Integrate function calling seamlessly into natural dialogue
• Provide Helpful Responses: Transform function results into meaningful user communications

Technical Implementation

The dataset follows industry-standard formats for function calling:

• OpenAI-compatible function schemas
• Structured JSON for function definitions and calls
• Clear separation between system instructions, user queries, and function responses
• Consistent formatting across all examples

This comprehensive training data ensures the model can handle real-world function calling scenarios with high accuracy and reliability, making it suitable for production deployment in AI assistant applications, automation workflows, and API integration tasks.

Technical Specifications

• Framework: Built using LLaMA-Factory
• Hardware Requirements: Recommended 80GB+ VRAM for inference
• Quantization: Compatible with various quantization methods (GPTQ, AWQ, etc.)
• Deployment: Suitable for both cloud and on-premise deployment

Citation

If you use this model in your research or applications, please cite:

@misc{qwen25-coder-glaive-toolcall,
  title={Qwen2.5-Coder-32B-Glaive-ToolCall},
  author={[RekklesAI]},
  year={2025},
  note={Fine-tuned version of Qwen2.5-Coder-32B-Instruct with enhanced tool calling capabilities using Glaive dataset}
}

License

apache-2.0

Acknowledgments

• Qwen Team: For the excellent base model Qwen2.5-Coder-32B-Instruct
• Glaive: For providing the high-quality tool calling dataset
• LLaMA-Factory: For the efficient fine-tuning framework

---

This model card follows the guidelines for responsible AI model documentation and transparency.