sheikh-bangla-110m

sheikh-bangla-110m

A Bengali GPT-2 language model trained from scratch with custom BPE tokenizer.

Model Details

• Model Type: GPT-2 (Causal Language Model)
• Language: Bengali (বাংলা)
• Training: Trained from scratch (no pretrained weights)
• Parameters: 23,142,400 (110M approx)
• Architecture: 6 layers, 8 attention heads, 512 hidden size

Tokenizer

• Type: BPE (Byte Pair Encoding)
• Vocabulary Size: 8,000 tokens
• Special Tokens:
  • [PAD] - Padding token (ID: 0)
  • [UNK] - Unknown token (ID: 1)
  • [CLS] - Beginning of sentence (ID: 2)
  • [SEP] - End of sentence (ID: 3)
  • [MASK] - Mask token (ID: 4)

Usage

from transformers import GPT2LMHeadModel, PreTrainedTokenizerFast

# Load model and tokenizer
model = GPT2LMHeadModel.from_pretrained("OsamaBinLikhon/sheikh-bangla-110m")
tokenizer = PreTrainedTokenizerFast.from_pretrained("OsamaBinLikhon/sheikh-bangla-110m")

# Generate text
input_text = "বাংলা ভাষা হলো"
inputs = tokenizer.encode(input_text, return_tensors="pt")
outputs = model.generate(inputs, max_length=100, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

• Training Data: Bengali text corpus
• Epochs: 20
• Batch Size: 4
• Learning Rate: 0.001
• Max Length: 256 tokens
• Optimizer: AdamW

Performance

The model shows progressive loss reduction during training:

• Initial Loss: ~8.08
• Final Loss: ~6.69 (after 20 epochs)

Limitations

• This is a small model trained on limited data
• May produce repetitive or nonsensical text for complex prompts
• Not suitable for production use without fine-tuning
• Training was done on CPU with limited computational resources

Citation

@misc{sheikh-bangla-110m,
  author = {Osama Bin Likhon},
  title = {sheikh-bangla-110m: Bengali GPT-2 Model Trained from Scratch},
  url = {https://huggingface.co/OsamaBinLikhon/sheikh-bangla-110m},
}

Acknowledgments

• Hugging Face for the Transformers library
• The tokenizers library for BPE implementation