Aparna852/de-en-translator-2

final-de-en-iwslt-model 🚀

This is a German to English translation model, fine-tuned over multiple stages starting from Helsinki-NLP/opus-mt-de-en.

✅ Training Stages

1. Base model: Helsinki-NLP/opus-mt-de-en
2. Stage 1 Dataset: wmt16 (German-English)
3. Stage 2 Dataset: Filtered wmt16 with better train/val split
4. Stage 3 Dataset: iwslt2017 (clean conversational corpus)

---

license: apache-2.0 tags: - translation - german - english - seq2seq - transformers - evaluation datasets: - iwslt2017 language: - de - en metrics: - sacrebleu - rouge - bertscore

🇩🇪➡️🇬🇧 de-en-translator-3

A transformer-based German → English translation model fine-tuned on the IWSLT2017 dataset using Hugging Face's Seq2SeqTrainer.

---

🚀 Model Overview

• ✅ Architecture: Seq2Seq (e.g., mBART / BART-style)
• 🔤 Direction: German → English
• 🧠 Trained using Hugging Face Transformers
• 🎯 Optimized with early stopping and BLEU-based evaluation
• 📦 Available on Hugging Face Hub for direct loading

---

📊 Evaluation Results

Tested on the IWSLT2017 test split:

Metric	Score
🔵 BLEU	39.23
🟢 ROUGE-L	0.67
🟣 BERTScore (F1)	0.9535

---

⚙️ Training Hyperparameters

Parameter	Value
Model Checkpoint	Aparna852/de-en-translator
Dataset	iwslt2017 (German-English)
Epochs	3
Train Batch Size	4
Eval Batch Size	4
Gradient Accumulation	8
Learning Rate	2e-5
Weight Decay	0.01
Warmup Steps	500
Max Sequence Length	128
FP16 (Mixed Precision)	True (if CUDA available)
Evaluation Strategy	epoch
Save Strategy	epoch
Logging Strategy	steps (every 10 steps)
Scheduler	linear
Metric for Best Model	eval_loss
Early Stopping	patience=2
Load Best Model at End	True
Trainer API	Seq2SeqTrainer from 🤗 Transformers

---

📥 Usage Example (Python)

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = AutoModelForSeq2SeqLM.from_pretrained("Aparna852/de-en-translator-3")
tokenizer = AutoTokenizer.from_pretrained("Aparna852/de-en-translator-3")

input_text = "Guten Morgen, wie geht es dir?"
inputs = tokenizer(input_text, return_tensors="pt")
output = model.generate(**inputs, max_length=128)
print(tokenizer.decode(output[0], skip_special_tokens=True))