Qwen2.5-3B-Instruct-math-aggressive

🎯 MATH-optimized | 📦 Aggressive pruning | ⚡ 25% weights pruned

This model is a aggressively pruned version of Qwen/Qwen2.5-3B-Instruct, specialized for MATH tasks using activation-aware weight pruning (Wanda-style).

✨ Key Features

• Specialization: Optimized for Math tasks
• Pruning Method: Wanda-style (|W| × |activation|) importance scoring
• Size Reduction: 25% weights pruned
• Use Case: Maximum compression for edge deployment

📊 Performance Comparison

Category	Original	Pruned	Change
Python	100.0%	20.0%	↓ 80.0%
Html	6.7%	6.7%	→
Trivia	66.7%	53.3%	↓ 13.3%
Math	60.0%	53.3% ⭐	↓ 6.7%
Reasoning	100.0%	100.0%	→
Medical	86.7%	60.0%	↓ 26.7%
Linux	100.0%	100.0%	→
Writing	73.3%	60.0%	↓ 13.3%

Average: 74.2% → 56.7% (-17.5%)

Math Retention: 88.9% of original performance

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CompactAI/Qwen2.5-3B-Instruct-math-aggressive")
tokenizer = AutoTokenizer.from_pretrained("CompactAI/Qwen2.5-3B-Instruct-math-aggressive")

# Example usage
inputs = tokenizer("Your prompt here", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📋 Technical Details

Property	Value
Base Model	Qwen/Qwen2.5-3B-Instruct
Specialization	Math
Prune Mode	Aggressive
Pruning Method	Activation-based weight pruning (Wanda)
Weight Reduction	25% weights pruned

🔗 Related Models

This model is part of the Qwen2.5-3B-Instruct pruned model collection. Variants:

• Safe - Conservative pruning (~10-20%), high accuracy retention
• Aggressive - Maximum compression (~40-50%), best for edge deployment

📜 License

This model inherits the license from the base model Qwen/Qwen2.5-3B-Instruct.

---

Generated by ZANNPS [Zeto Automatic Neural Network Pruning System]