Model Card for aktheroy/FT_Translate_en_el_hi

This model is a fine-tuned version of facebook/m2m100_418M, designed for multilingual translation tasks between English (en), Greek (el), and Hindi (hi). The model achieves efficient translation by leveraging the M2M100 architecture, which supports many-to-many language translation.

Model Details

Model Description

• Developed by: Aktheroy
• Model type: Transformer-based encoder-decoder
• Language(s) (NLP): English, Hindi, Greek
• License: MIT
• Finetuned from model: facebook/m2m100_418M

Model Sources

• Repository: aktheroy/FT_Translate_en_el_hi

Uses

Direct Use

The model can be used for translation tasks between the supported languages (English, Hindi, Greek). Use cases include:

• Cross-lingual communication
• Multilingual content generation
• Language learning assistance

Downstream Use

The model can be fine-tuned further for domain-specific translation tasks, such as medical or legal translations.

Out-of-Scope Use

The model is not suitable for:

• Translating unsupported languages
• Generating content for sensitive or harmful purposes

Bias, Risks, and Limitations

While the model supports multilingual translations, it might exhibit:

• Biases from the pretraining and fine-tuning datasets.
• Reduced performance for idiomatic expressions or cultural nuances.

Recommendations

Users should:

• Verify translations, especially for critical applications.
• Use supplementary tools to validate outputs in sensitive scenarios.

How to Get Started with the Model

Here is an example of how to use the model for translation tasks:

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_name = "aktheroy/FT_Translate_en_el_hi"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Example input
input_text = "Hello, how are you?"
tokenizer.src_lang = "en"
tokenizer.tgt_lang = "hi"

# Tokenize and generate output
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)
print(translation)

Training Details

Training Data

The model was fine-tuned on a custom dataset containing parallel translations between English, Hindi, and Greek.

Training Procedure

Preprocessing

The dataset was preprocessed to:

• Normalize text.
• Tokenize using the M2M100 tokenizer.

Training Hyperparameters

• Epochs: 10
• Batch size: 16
• Learning rate: 5e-5
• Mixed Precision: Disabled (FP32 used)

Speeds, Sizes, Times

• Training runtime: 20.3 hours
• Training samples per second: 17.508
• Training steps per second: 0.137
• Final training loss: 0.873

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated on a held-out test set from the same domains as the training data.

Metrics

• BLEU score (to be computed during final evaluation).

Results

• Training Loss: 0.873
• Detailed BLEU score results will be provided in subsequent updates.

Environmental Impact

• Hardware Type: MacBook with M3 Pro
• Hours used: 20.3 hours
• Cloud Provider: Local hardware
• Carbon Emitted: Minimal (local training)

Technical Specifications

Model Architecture and Objective

The model is based on the M2M100 architecture, a transformer-based encoder-decoder model designed for multilingual translation without relying on English as an intermediary language.

Compute Infrastructure

Hardware

• Device: MacBook with M3 Pro

Software

• Transformers library from Hugging Face
• Python 3.12

Citation

If you use this model, please cite it as:

APA: Aktheroy (2025). Fine-Tuned M2M100 Translation Model. Hugging Face. Retrieved from https://huggingface.co/aktheroy/FT_Translate_en_el_hi

Model Card Authors

• Aktheroy

Model Card Contact

For questions or feedback, contact the author via Hugging Face.