This area progresses from token representations and recurrent classifiers to sequence-to-sequence systems and a full local language-model project. Each numbered source-backed folder is intentionally small enough to inspect; configuration and data paths are local rather than globally packaged.
| Area | Implemented source-backed families | Documentation |
|---|---|---|
| Word representations | One-hot encoding, co-occurrence, PPMI-SVD, CBOW, Skip-Gram, GloVe, FastText | Open |
| Sequence classification | Vanilla RNN, LSTM, GRU, bidirectional RNN, stacked RNN | Open |
| Seq2Seq | Basic encoder-decoder, attention, copy mechanism, coverage, Transformer seq2seq | Open |
| End-to-end project | Reddit scraping, corpus preparation, BPE, TFRecords, decoder-only Transformer, evaluation, sampling | RedditStory |
AttentionMechanisms/, DiffusionText/, GenerativeModels/,
PretrainedLanguageModels/, and Transformers/ are present as numbered
learning-roadmap directories but contain no tracked Python implementation files
in this checkout. They are intentionally not described as implemented models.
The source uses tf.keras.Model and custom layers for sequence and attention
architectures; text input paths use tf.data where a dataset implementation is
present. Several trainers use tf.distribute.MirroredStrategy, while the
RedditStory project additionally has a TFRecord pipeline, explicit causal mask
logic, GradientTape training, and optional XLA controls.
The sequence-model and seq2seq dataset base classes contain explicit
NotImplementedError paths for adapters. Their pages distinguish the model
architecture from the incomplete dataset wiring, and no benchmark result is
claimed without a retained report.