Sequence Model Examples

Runnable sequence-model examples, organized by workflow.

Main entrypoints:

• CharGpt (torchlean chargpt): Karpathy-style char-level GPT on a single text file (Tiny Shakespeare). This is the teaching path for character tokenization, with separate smoke and lecture presets.
• Gpt2 (torchlean gpt2): byte-level GPT-2-style causal Transformer with a small, local-friendly config. Use this when you want to see masked self-attention + LayerNorm + FFN wiring, and a save/reload path via Gpt2Saved.
• TextGpt2 (torchlean text_gpt2): CUDA-only bounded corpus trainer with byte-level or GPT-2 BPE tokenization. Its BPE path projects observed token IDs into a local output vocabulary.
• Mamba (torchlean mamba): compact text walkthrough for the Mamba-style model.

Other sequence examples:

• Rnn and Lstm: compact real-text recurrent training checks over the shared corpus-data boundary.
• Transformer: one-block encoder example for attention/norm/FFN wiring.
• GptAdder: synthetic algorithmic curriculum (addition), runnable as torchlean gpt_adder.

For supervised time-series forecasting with an LSTM, see NN.Examples.Models.Supervised.LstmRegression (torchlean lstm_regression).