Model Card for oopere/pruned40-gemma-2-2b

This model is a pruned version of the Gemma-2b architecture, with a parameter reduction of 40% in the MLP Layers. The pruning process aims to enhance computational efficiency while maintaining acceptable performance across specific tasks. This model is not intended to be used directly, but rather to be fine-tuned for specific tasks where it can achieve equal or superior performance compared to fine-tuning the base model for the same task.

Model Details

• Model Type: Pruned version of Gemma-2b using structured pruning
• Original Model: google/gemma-2-2b
• Pruning Method: Structured pruning of MLP layers using importance scores based on absolute maximum weights
• Size Reduction: 11.36% (from 2.2B to 1.95B parameters)
• Architecture: Same as original Gemma but with reduced MLP layer sizes
• Language(s): Same as original model
• License: Gemma
• Developed by: Pere Martra

Key Findings

• Maintains moderate performance on binary classification tasks (BoolQ)
• Significant but manageable degradation on reasoning tasks (ARC-Easy)
• Substantial impact on long-range comprehension (LAMBADA)
• Notable increase in perplexity (from 3.71 to 29.68 on LAMBADA-OpenAI)

Limitations

• Considerable reduction in performance on complex language understanding tasks
• Significant degradation in long-range dependency handling
• May not be suitable for applications requiring high accuracy on language completion tasks
• Best suited for simpler classification tasks

Implementation Details

• Pruning Notebook: Detailed implementation and methodology
• GitHub Repository: LLM Course

Pruning Method

• Technique: Structured pruning targeting MLP layers
• Pruning Ratio: 40% of neurons removed from MLP layers
• Selection Criteria: Importance scoring based on absolute maximum weights
• Architecture Specifics: Maintained original architecture structure during pruning

Hardware Requirements

Memory Requirements

• Base Model:

  • Parameters: ~4.4 GB (FP16)
  • Total Runtime Memory: ~5.5 GB

• Pruned Model (40%):

  • Parameters: ~3.9 GB (FP16)
  • Total Runtime Memory: ~4.9 GB

• Memory Reduction:

  • Parameter Memory: 11.36%
  • Total Runtime Memory: ~10.9%

Notes:

• Memory requirements assume FP16 precision
• Actual memory usage may vary depending on:
  • Batch size
  • Sequence length
  • Implementation details
  • Runtime environment

Minimum Requirements

• GPU Memory: 6GB for base model, 5GB for pruned model
• CPU Memory: 16GB recommended for both models

Acknowledgments

• Thanks to Mariusz Kurman for creating llama-pruning, a library that implements and extends this pruning methodology.