jeffmeloy's picture
Update README.md
f971ac5 verified
|
raw
history blame
2.26 kB
metadata
license: apache-2.0
base_model:
  - Qwen/Qwen2.5-7B
pipeline_tag: text-generation
tags:
  - not-for-all-audiences
language:
  - en
library_name: transformers

Model Description

Model created by analyzing and selecting the optimal layers from other Qwen2.5-7B models based on their dimensional utilization efficiency, measured by the Normalized Effective Rank (NER). Computed like:

Singular Value Decomposition:

  • Input: Weight matrix A ∈ R^(m×n) # m = number of output features, n = number of input features
  • Compute singular values σᵢ where σᵢ ≥ 0 # σᵢ represents the importance of each dimension
  • Filter values above numerical threshold (>1e-12) # removes numerical noise from computation

Distribution Normalization:

  • Sum all singular values: S = Σσᵢ # S acts as normalization factor
  • Create probability distribution: pᵢ = σᵢ/S # converts singular values to probabilities summing to 1

Entropy Calculation:

  • Compute Shannon entropy: H = -Σ(pᵢ * log₂(pᵢ)) # measures information content of distribution
  • Calculate maximum possible entropy: H_max = log₂(n) # n = number of singular values where n is the number of singular values # maximum entropy occurs when all dimensions contribute equally

Normalization:

  • Final NER score = H/H_max # normalizes score to [0,1] range
  • Results in value between 0 and 1 # 0 = single dimension dominance, 1 = perfect dimensional utilization
  • Higher scores indicate more uniform dimensional utilization

Creating Composite Model

Layer Analysis:

  • Download base and fine-tuned models from Hugging Face Hub # fetches models using Hugging Face API
  • Calculate Normalized Effective Rank (NER) for each layer within each model # process each independently

Layer Selection:

  • Identify common layer structures across models
  • Define model and layer name pairs that have highest NER for each layer based on their NER scores

Model Composition:

  • Incrementally build a composite model using layer with highest NER from model pool.

Output Generation:

  • Save merge reports documenting layer sources
  • Copy config and tokenizer files from base model
  • Save the composite model with complete weights # model ready to use