Knowledge encoding by examples of Word2Vec and LLM training

This repository contains weights for a list of language models:

word2vec.pt: embedding trained on 150mil pairs of text tokens subsampled from text8 dataset. SkipGram method with negative sampling was used as described in the original paper.
mlp.pt: 2-layers MLP trained on the same dataset and using pretrained embeddings.
mlp_norm.pt: Version of the MLP model utilizing LayerNorm for better scaling of the learned features distribution.