word2vec (Skip-gram + negative sampling)

Small NumPy-only trainer: builds a vocabulary from plain text, samples Skip-gram (center, context) pairs with a random window, and learns separate input/output embedding matrices with negative sampling.

Setup

• Python 3.10+
• numpy

Data

Point training at a UTF-8 text file. The script defaults to text8.txt (e.g. the text8 corpus). Place that file in the project directory or pass another path via run_training(text_path=...).

Run

python skipgram_neg_sampleing.py

main() calls run_training() with a small max_chars and more epochs for a quick demo. For full control (window size, batch size, negative samples, etc.), import run_training from skipgram_neg_sampleing or extend the script.

Layout

File	Role
dataset.py	Tokenize text, vocab + counts, negative-sampling distribution, Skip-gram pairs
skipgram_neg_sampleing.py	Loss, gradients, SGD training loop

Trained input embeddings are the usual word vectors (input_emb returned from run_training).