Logo Lanfrica

ianMuchesia/tiny-translation-model

Domaine:

natural language processing

Type de record:

softwaremodel
Créateur:
ian
Hôte:
Minimal seq2seq Transformer for English->Swahili translation with attention-based encoder/decoder blocks, tokenization, and dataset utilities. Tiny Translation Model ====================== A compact seq2seq Transformer for English->Swahili translation in PyTorch. ## Features - **Encoder/Decoder Blocks** (`src/encoder.py`, `src/decoder.py`): Multi-head attention and feed-forward layers - **Positional Encoding** (`src/positional_encoding.py`): Sinusoidal position signals - **Tokenizer and Vocab** (`src/tokenizer.py`): Word-level tokenization with special tokens - **Dataset Loader** (`src/dataset.py`): Parses "english => swahili" pairs from `data/toy_translations.txt` - **Seq2Seq Wrapper** (`src/encoder_decoder.py`): End-to-end encoder-decoder model - **Training Entry Point** (`src/train.py`): Training script (in progress) - **Training Notebook** (`notebooks/transation_examples.ipynb`): End-to-end setup and training loop - **Attention Debugging**: Heatmaps and shape checks for masking and decoder behavior - **Inference Module** (`src/inference.py`): Decoding and translation utilities - **BLEU Evaluation** (`src/evaluate_bleu.py`): Automatic translation quality assessment - **Greedy Decoding** (`src/greedy_decode.py`): Fast single-path inference strategy - **Test Dataset** (`data/toy_translations_test.txt`): Holdout evaluation set ## Quick Start ```bash python -m src.train ``` ## Status - Training, evaluation, and inference pipelines are complete and working. - Masking, accuracy calculations, and attention heatmap analysis validated. - BLEU scores and sample predictions generated and saved. ## Project Structure - `src/` — Core implementation - `data/` — Translation pairs - `notebooks/` — Experiments and demos - `experiments/` — Saved outputs - `math-notes/` — Derivations and notes

Languages