SemiQwenn-1.5b

SemiQwenn - Distilled Qwen2.5 1.5B

Model Description

SemiQwenn-1.5B is a distilled version of Devstral knowledge transferred to the efficient Qwen2.5-1.5B architecture through Supervised Fine-Tuning (SFT) distillation. The model learns from Devstral's responses on the training dataset, effectively capturing the teacher model's capabilities while maintaining the computational efficiency of the Qwen2.5-1.5B architecture. This model was created as part of a datathon project focused on efficient language model training and deployment.

Model Details

• Model Name: SemiQwenn-1.5B
• Student Model: Qwen2.5-1.5B
• Teacher Model: Devstral
• Model Size: 1.5 billion parameters
• Training Method: SFT (Supervised Fine-Tuning) Distillation with LoRA adapters
• Language(s): English (primary), with multilingual capabilities inherited from base model
• License: Same as base Qwen2.5 model
• Model Type: Causal Language Model

Training Details

Training Data

• Dataset: Code Alpaca + GSM8K (30k samples)
• Training Split: Stratified split for balanced learning
• Data Format: JSONL format with instruction-response pairs

Training Configuration

• Training Method: LoRA (Low-Rank Adaptation)
• Teacher Model: Devstral (for SFT distillation)
• Training Framework: Transformers/PEFT
• Hardware: GPU-optimized training

Training Process

• Fine-tuned using SFT distillation from Devstral (teacher) to Qwen2.5-1.5B (student)
• LoRA adapters applied to the student model and merged for final deployment
• Optimized to transfer Devstral's knowledge to the more efficient Qwen architecture

Performance

GSM8K Evaluation Results

• Model demonstrates competitive performance on mathematical reasoning tasks
• Evaluation results available in project evaluation files
• Comparison with base models and teacher model included

Resource Usage

• Balanced trade-off between performance and efficiency
• More capable than 0.5B version while remaining resource-efficient
• Optimized for deployment scenarios requiring higher performance

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model
model = AutoModelForCausalLM.from_pretrained("alfiwillianz/SemiQwenn-1.5b")
tokenizer = AutoTokenizer.from_pretrained("alfiwillianz/SemiQwenn-1.5b")

# Example usage
prompt = "Solve this math problem: What is 15 * 24?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Model Architecture

• Architecture: Transformer-based decoder-only model
• Attention: Multi-head attention mechanism
• Vocabulary Size: Inherited from Qwen2.5 tokenizer
• Context Length: Supports extended context as per base model

Intended Use

Primary Use Cases

• Mathematical reasoning and problem solving
• Code generation and understanding
• Educational applications
• Research in efficient language models
• Applications requiring better performance than 0.5B models

Out-of-Scope Uses

• This model should not be used for generating harmful, biased, or inappropriate content
• Not suitable for high-stakes decision making without human oversight
• Not designed for real-time critical applications

Limitations and Biases

• As a 1.5B parameter model, it has limitations compared to larger models
• May inherit biases from training data and base model
• Performance may vary on tasks outside the training distribution
• Limited by the knowledge cutoff of the base model

Ethical Considerations

• Model outputs should be reviewed for accuracy, especially in educational contexts
• Users should be aware of potential biases and limitations
• Appropriate safeguards should be implemented for production use

Citation

If you use SemiQwenn-1.5B in your research or applications, please cite:

@misc{semiqwenn1.5b2025,
  title={SemiQwenn-1.5B: A Distilled Qwen2.5 1.5B Model},
  author={Alfi Willianz},
  year={2025},
  note={Knowledge distilled model based on Qwen2.5-1.5B}
}

Acknowledgments

• Built upon Qwen2.5 by Alibaba Cloud
• Training methodology inspired by knowledge distillation techniques
• Part of Datathon 2025 project on efficient language models

Model Files

This directory contains:

• Merged model weights combining LoRA adapters with base model
• Tokenizer configuration
• Model configuration files
• Training artifacts and logs

Contact

For questions about this model or the training process, please refer to the project documentation or contact the development team.

---