BenChaliah/Superposition-Transformer

a novel architecture that leverages Autoencoders to superimpose the hidden representations of a base model and a fine-tuned model within a shared parameter space. Using B-spline-based blending coefficients and autoencoders that adaptively reconstruct the original hidden states based on the input data distribution.

25
/ 100
Experimental

This project helps machine learning engineers and researchers adapt large language models (LLMs) to new tasks or languages without losing their original abilities. It takes an existing LLM and a version fine-tuned for a specific task or domain, combining their knowledge to produce a single, more versatile model. This is for AI practitioners who need to deploy models that can handle multiple specialized tasks while retaining general knowledge.

No commits in the last 6 months.

Use this if you need to fine-tune a large language model for a new task or domain and want to avoid the problem of 'catastrophic forgetting,' where the model loses its previous knowledge.

Not ideal if you are looking for a pre-trained model ready for immediate deployment or if you are not comfortable working with deep learning architectures and model training.

large-language-models model-adaptation multi-task-learning catastrophic-forgetting natural-language-processing
No License Stale 6m No Package No Dependents
Maintenance 2 / 25
Adoption 9 / 25
Maturity 8 / 25
Community 6 / 25

How are scores calculated?

Stars

79

Forks

3

Language

Jupyter Notebook

License

Last pushed

Aug 01, 2025

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/transformers/BenChaliah/Superposition-Transformer"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.