Skip to content

Latest commit

 

History

History
153 lines (125 loc) · 6.63 KB

README.md

File metadata and controls

153 lines (125 loc) · 6.63 KB

Fine-Tuned English-to-Japanese Translation Model

This repository contains a fine-tuned version of the Helsinki-NLP/opus-mt-en-jap model for English-to-Japanese translation, trained using the KDE4 dataset.

The fine-tuning process enhances translation quality, specifically for domain-specific sentences found in KDE software documentation and similar contexts.

Features

  • Enhanced English-to-Japanese translation for technical and software-related contexts.
  • Fine-tuned using the KDE4 dataset for improved domain relevance.

Model

The fine-tuned model is hosted on Hugging Face: sattu-finetuned-kde4-en-to-jap.

How to Use

You can load the model directly from Hugging Face:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("iSathyam03/sattu-finetuned-kde4-en-to-jap")
model = AutoModelForSeq2SeqLM.from_pretrained("iSathyam03/sattu-finetuned-kde4-en-to-jap")

# Translate text
text = "This is a sample English sentence."
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
outputs = model.generate(**inputs)
translated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translated_text)

Training Process

Dataset

  • KDE4 Dataset: Contains parallel English-Japanese sentences from KDE software documentation. This dataset ensures the translations are relevant to technical and domain-specific contexts.

Training Arguments

The model was fine-tuned using the following Seq2SeqTrainingArguments:

from transformers import Seq2SeqTrainingArguments

args = Seq2SeqTrainingArguments(
    f"sattu-finetuned-kde4-en-to-jap",
    evaluation_strategy="no",
    save_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=16,
    weight_decay=0.01,
    save_total_limit=3,
    num_train_epochs=3,
    predict_with_generate=True,
    fp16=True,
    push_to_hub=True,
)

Framework

Base Model: Helsinki-NLP/opus-mt-en-jap. Libraries: Transformers and PyTorch.

Evaluation Metrics

The fine-tuning process significantly improved the model's performance, as shown by the evaluation metrics:

Metric Before Fine-Tuning After Fine-Tuning
Evaluation Loss 10.38 1.42
BLEU Score 0.0046 20.73
Evaluation Runtime (s) 522.76 503.29
Samples per Second 25.14 26.11
Steps per Second 1.57 1.63
  • BLEU Score: Improved from near-zero (0.0046) to 20.73, reflecting the enhanced translation quality for English-to-Japanese.
  • Evaluation Loss: Reduced from 10.38 to 1.42, showing the model's increased accuracy.
  • Other metrics, such as runtime and throughput, show slight improvements due to fine-tuning optimizations.

Constraints and Recommendations

This project was developed using limited hardware resources, specifically an RTX 4050 GPU with 6GB of VRAM. While this setup worked for fine-tuning, users may face constraints when working with larger models or datasets.

If you're using limited resources, consider the following recommendations to optimize your training process. Alternatively, you can use platforms like Google Colab or more powerful hardware setups for a smoother experience.

Tips for Working with Limited Resources

  1. Reduce per_device_train_batch_size: A large batch size can exceed your GPU's memory capacity. Reduce it as needed.
per_device_train_batch_size=8  # Reduce to 16 or 8 if needed
  1. Reduce per_device_eval_batch_size: Similarly, reduce the evaluation batch size to save memory.
per_device_eval_batch_size=16  # Reduce to 16 or 8
  1. Use gradient_accumulation_steps: Simulate a larger batch size by accumulating gradients across steps
gradient_accumulation_steps=2  # Simulate larger batch size
  1. Enable Mixed Precision Training (fp16): Mixed precision training reduces memory usage by using 16-bit floating-point precision
fp16=True  # Already enabled in this project
  1. Remove Unused Columns and Efficient Padding: If your dataset contains unused columns, remove them to save memory. Set label_pad_token_id for text generation tasks.
remove_unused_columns=True
label_pad_token_id=-100  # For text generation or seq-to-seq tasks
  1. Use save_steps instead of save_strategy="epoch": If memory usage is a concern, saving the model after every epoch can lead to frequent storage usage. You can save more frequently using save_steps.
save_steps=500  # Save every 500 steps or as needed
  1. Limit push_to_hub: Pushing to the hub might be memory-intensive. If you're not actively pushing to the hub during training, consider disabling it until you're ready.
push_to_hub=False  # Turn off unless you need to push to Hugging Face Hub during training

Example Updated Code:

from transformers import Seq2SeqTrainingArguments

args = Seq2SeqTrainingArguments(
    output_dir="sattu-finetuned-kde4-en-to-jap",
    evaluation_strategy="no",
    save_strategy="epoch",  # Or save_steps=500 if you prefer more frequent saving
    learning_rate=2e-5,
    per_device_train_batch_size=16,  # Reduced batch size
    per_device_eval_batch_size=32,  # Reduced evaluation batch size
    weight_decay=0.01,
    save_total_limit=3,
    num_train_epochs=3,
    predict_with_generate=True,
    fp16=True,  # Mixed precision training enabled
    push_to_hub=False,  # Turned off unless needed
    gradient_accumulation_steps=2,  # Accumulate gradients for larger effective batch size
    remove_unused_columns=True,  # Remove unused columns to save memory
    label_pad_token_id=-100  # Used in text generation tasks
)

The Jupyter Notebook detailing the training process is included: code.ipynb

Dependencies

Install the required dependencies using:

pip install -r requirements.txt

Acknowledgements

  • Hugging Face and the Transformers library.
  • KDE4 dataset for domain-specific parallel data.
  • Helsinki-NLP for the base translation model.

License

This project is licensed under the MIT License. See the LICENSE file for details.