aijadugar/multi30k-lstm

Seq2Seq LSTM Translator (English -> German)

A PyTorch Seq2Seq Encoder-Decoder LSTM trained from scratch on the Multi30k dataset.

Model Details

• Architecture: Seq2Seq
• Framework: PyTorch
• Dataset: Multi30k (English -> German)
• Embedding Dimension: 128
• Hidden Dimension: 256

Training

• Optimizer: Adam
• Learning Rate: 1e-3
• Loss: CrossEntropyLoss (ignore_index=PAD)
• Gradient Clipping: max_norm=1.0
• Epochs: 50
• Batch Size: 64

Dataset

Split	Samples
Train	29,000
Validation	1,014
Test	1,000

Performance

Metric	Value
Train Loss	0.62
Validation Loss	5.55
Test Loss	3.79
BLEU	14.70

Usage

# Load model
model = Seq2Seq(encoder, decoder, device)

model.load_state_dict(
    load_file("best_seq2seq_lstm.safetensors")
)

model.eval()

# Inference
sentence = "A man is riding a bicycle."

translation = translate_sentence(
    sentence,
    model
)

print(" ".join(translation))

Author Ankit Bari

GitHub: https://github.com/aijadugar Hugging Face: https://huggingface.co/aijadugar