FastText_LMM_hack / README.md
ccore's picture
Update README.md
e45c53b
metadata
license: mit
pipeline_tag: text-generation

Advanced FastText Modification

colab:https://colab.research.google.com/drive/1apd0UFutEqBMG_7t_60QJAoxWxNdc1yr#scrollTo=GayYJ9VeJms7

Overview

This software is an innovative version of FastText, optimized to use its classification system for fast and efficient word prediction. With capabilities that can surpass modern language models, it delivers exceptional performance.

Model Training

  • Data Used: The models were trained with more than 20,000 question and answer (QA) interactions from the Wizard Evo instruct dataset.
  • Focus on Dynamic Learning: This approach guarantees in-depth understanding in contexts of dialogue and interaction.

Processing Capacity

  • Speed: 800 tokens/thread/second.
  • Performance on Common Hardware: On a notebook with 16 threads, the program processes approximately 50 billion tokens per hour.

Learning and Testing

  • Learning: Evidence points to an ability to learn robustly.
  • Testing: The software is still in the early testing phases, but preliminary results are promising.

This project is constantly evolving and we welcome any feedback or contributions from the community!