Xlmr Large Qa Fa

🚀 XLM-RoBERTa large for QA (PersianQA - 🇮🇷)

This model is a fine - tuned version of [xlm - roberta - large](https://huggingface.co/xlm - roberta - large) on the PersianQA dataset. It is designed for question - answering tasks, supporting both Persian and multiple languages.

✨ Features

• Multilingual Support: Supports both Persian (fa) and multiple languages.
• Question - Answering Task: Specifically fine - tuned for question - answering tasks on the PersianQA dataset.

📦 Installation

The installation steps are not provided in the original README, so this section is skipped.

💻 Usage Examples

Basic Usage

from transformers import AutoModelForQuestionAnswering, AutoTokenizer, pipeline

model_name_or_path = "m3hrdadfi/xlmr-large-qa-fa"
nlp = pipeline('question-answering', model=model_name_or_path, tokenizer=model_name_or_path)

context = """
شب یَلدا یا شب چلّه یکی از کهن‌ترین جشن‌های ایرانی است. 
در این جشن، طی شدن بلندترین شب سال و به دنبال آن بلندتر شدن طول روزها
 در نیم‌کرهٔ شمالی، که مصادف با انقلاب زمستانی است، گرامی داشته می‌شود. 
نام دیگر این شب «چِلّه» است، زیرا برگزاری این جشن، یک آیین ایرانی‌است.
"""
# Translation [EN]
# context = [
  # Yalda night or Cheleh night is one of the oldest Iranian celebrations. 
  # The festival celebrates the longest night of the year, followed by longer days in the Northern Hemisphere, 
  # which coincides with the Winter Revolution. 
  # Another name for this night is "Chelleh", because holding this celebration is an Iranian ritual.
# ]


questions = [
    "نام دیگر شب یلدا؟",
    "کهن ترین جشن ایرانی‌ها چه است؟",
    "شب یلدا مصادف با چه پدیده‌ای است؟"
]
# Translation [EN]
# questions = [
  # Another name for Yalda night?
  # What is the ancient tradition of Iranian celebration?
  # What phenomenon does Yalda night coincide with?
# ]


kwargs = {}

for question in questions:
    r = nlp(question=question, context=context, **kwargs)
    answer = " ".join([token.strip() for token in r["answer"].strip().split() if token.strip()])
    print(f"{question} {answer}")

Advanced Usage

The original README does not provide advanced usage examples, so this part is skipped.

📚 Documentation

Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e - 05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 20.0
• mixed_precision_training: Native AMP

Performance

Evaluation results on the eval set with the official eval script.

Evalset

"HasAns_exact": 58.678955453149,
"HasAns_f1": 82.3746683591845,
"HasAns_total": 651,
"NoAns_exact": 86.02150537634408,
"NoAns_f1": 86.02150537634408,
"NoAns_total": 279,
"exact": 66.88172043010752,
"f1": 83.46871946433232,
"total": 930

🔧 Technical Details

The model is based on the XLM - RoBERTa large architecture and is fine - tuned on the PersianQA dataset for question - answering tasks.

📄 License

The original README does not provide license information, so this section is skipped.

👥 Authors

• Mehrdad Farahani

🛠️ Framework versions

• Transformers 4.12.0.dev0
• Pytorch 1.9.1+cu111
• Datasets 1.12.1
• Tokenizers 0.10.3