VyasVani: Sanskrit LSTM Language Model

Hosted on Hugging Face Hub: https://huggingface.co/Henil1/VyasVani

Model Description

VyasVani is a character-level Sanskrit text generation model based on a 4-layer LSTM. It was trained on a corpus of classical Sanskrit ślokas (verses) to learn patterns of Sanskrit meter and grammar. Ideal for generating coherent Sanskrit text and poetic verses.

• Architecture: 4-layer LSTM
• Embedding Size: 128
• Hidden Size: 256
• Vocabulary Size: Auto-detected from the training corpus
• Total Parameters: 2,147,907 (all trainable)

Usage

import json
import torch
from huggingface_hub import hf_hub_download
from example_inference import SanskritLSTM, generate_text

# Download files from the Hub
model_path = hf_hub_download(repo_id="Henil1/VyasVani", filename="pytorch_model.bin")
config_path = hf_hub_download(repo_id="Henil1/VyasVani", filename="config.json")
stoi_path   = hf_hub_download(repo_id="Henil1/VyasVani", filename="stoi.json")
itos_path   = hf_hub_download(repo_id="Henil1/VyasVani", filename="itos.json")

# Load configuration and tokenizer
config = json.load(open(config_path))
stoi   = json.load(open(stoi_path))
itos   = json.load(open(itos_path))

# Rebuild the model
model = SanskritLSTM(
    vocab_size=config["vocab_size"],
    embed_size=config.get("embed_size",128),
    hidden_size=config.get("hidden_size",256),
    num_layers=config.get("num_layers",4)
)
model.load_state_dict(torch.load(model_path, map_location="cpu"))
model.eval()

# Generate Sanskrit text
seed = "श्री"
generated = generate_text(model, seed, length=200, temperature=0.8, top_k=5)
print(generated)

Example Inference Script

See example_inference.py in the repository for a ready-to-run script defining SanskritLSTM and generate_text functions.

Evaluation Metrics

• Perplexity: 5.116

Training Details

• Dataset: Classical Sanskrit ślokas from various texts
• Training Steps: 1 Epoch
• Optimizer: Adam (lr=1e-3)
• Batch Size: 64
• Sequence Length: 128

Citation

If you use VyasVani in your work, please cite:

@misc{VyasVani2025,
  title={VyasVani: A 4-Layer LSTM for Sanskrit Text Generation},
  author={Henilsinh Raj},
  year={2025},
  howpublished={\url{https://huggingface.co/Henil1/VyasVani}}
}

License

This model is released under the MIT License. See LICENSE for details.