🇩🇪➡️🇬🇧 de-en-translator

A transformer-based German → English translation model fine-tuned on a custom split of the WMT16 (de-en) dataset using 🤗 Transformers and Seq2SeqTrainer.

---

🧠 Model Details

• ✅ Model: Aparna852/german-english-translator (fine-tuned)
• 🔤 Task: German ➡️ English Translation
• 📚 Dataset: WMT16 (wmt/wmt16 - de-en)
• ⚙️ Strategy: Custom train/val/test split, truncated input
• 🧪 Evaluation Metrics: BLEU (via sacrebleu)

---

⚙️ Training Hyperparameters

Parameter	Value
Dataset	wmt/wmt16 (German-English)
Train Size	~2.5% of original training set
Validation Size	~2.8% of original validation
Max Length	64
Epochs	3
Train Batch Size	4
Eval Batch Size	4
Gradient Accumulation	8
Learning Rate	1e-5
Weight Decay	0.03
Warmup Steps	500
FP16 (Mixed Precision)	True (if CUDA available)
Scheduler	linear
Evaluation Strategy	epoch
Save Strategy	epoch
Logging Steps	10
Early Stopping	patience=2
Metric for Best Model	eval_loss
Trainer API	Seq2SeqTrainer from 🤗 Transformers

---

📊 Evaluation Setup

You can run the evaluation after training using:

from evaluate import load
bleu = load("sacrebleu")

# Compute BLEU on tokenized test dataset
preds = [...]  # Generated translations
refs = [...]   # Reference translations
bleu.compute(predictions=preds, references=[[r] for r in refs])