README.md

metadata

license: apache-2.0
tags:
  - pruned
  - writing
  - optimized
  - wanda
  - activation-pruning
base_model: Qwen/Qwen3-0.6B
pipeline_tag: text-generation

Qwen3-0.6B-writing-safe

🎯 WRITING-optimized | 📦 Safe pruning | ⚡ 1% weights pruned

This model is a conservatively pruned version of Qwen/Qwen3-0.6B, specialized for WRITING tasks using activation-aware weight pruning (Wanda-style).

✨ Key Features

• Specialization: Optimized for Writing tasks
• Pruning Method: Wanda-style (|W| × |activation|) importance scoring
• Size Reduction: 1% weights pruned
• Use Case: High accuracy retention, ideal for production use

📊 Performance Comparison

Category	Original	Pruned	Change
Python	30.0%	30.0%	→
Html	0.0%	0.0%	→
Trivia	90.0%	90.0%	→
Math	96.7%	100.0%	↑ 3.3%
Reasoning	36.7%	36.7%	→
Medical	83.3%	83.3%	→
Linux	93.3%	93.3%	→
Writing	53.3%	56.7% ⭐	↑ 3.3%

Average: 60.4% → 61.2% (+0.8%)

Writing Retention: 106.2% of original performance

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CompactAI/Qwen3-0.6B-writing-safe")
tokenizer = AutoTokenizer.from_pretrained("CompactAI/Qwen3-0.6B-writing-safe")

# 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/Qwen3-0.6B
Specialization	Writing
Prune Mode	Safe
Pruning Method	Activation-based weight pruning (Wanda)
Weight Reduction	1% weights pruned

🔗 Related Models

This model is part of the Qwen3-0.6B 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/Qwen3-0.6B.

---

Generated by ZANNPS [Zeto Automatic Neural Network Pruning System]