mouadtarif/zero-padding-model-merging

Monolingual Hybridization Research

Author: Mouad Tarif
Date: April 2026
Type: Independent Research Log

Summary

This work documents experiments on hybridizing monolingual language models without retraining. The goal is to create a functional merged model that preserves both original languages and exhibits emergent cross-lingual capabilities.

Papers

Paper 1: Hybridizing the Incompatible

Date: 11 April 2026
DOI: 10.5281/zenodo.19520544

Experiment: Monolingual model + NLLB (Meta).
Result: ≈50% success, arithmetic collapse, NLLB dominance.
Insight: "Hybridizing is like genetic engineering — the host may reject foreign genes unless made compatible."

Paper 2: The Final Hybrid

Date: 14 April 2026
DOI: 10.5281/zenodo.19562409

Method: Zero-padding expansion for 2D tensors, repetition for 1D tensors. RAM limited to 4GB. Tokenizer alignment (tokenizer.json vs vocab.json).
Result: Both original languages preserved, no single model dominates, emergent cross-lingual capabilities observed.

Code

Primary Source: GitHub: Monolingual-Hybridization
Date: 14 April 2026

Note: The files below are re-published here for archival and citation purposes. The GitHub repository remains the primary source for the latest version and commit history.

Re-published Files

• hybridize_models.py — Main hybridization script
• load_and_test_hybrid.py — Load and test script

Original Date: 14 April 2026
Re-published in Hugging Face: [Date you upload]

Citation

@misc{tarif2026hybrid,
  title={Monolingual Hybridization Research},
  author={Tarif, Mouad},
  year={2026},
  doi={10.5281/zenodo.19520544}
}