mamba-tamil-xlit

Mamba Tamil Transliteration

A Mamba-based sequence-to-sequence model for English-to-Tamil transliteration.

Model Details

• Architecture: Mamba encoder-decoder with cross-attention
• Parameters: 6.5M
• Training data: Dakshina Tamil lexicon (train split)
• Tokenization: Grapheme-level

Performance

Beats IndicXlit (78.53%) on Dakshina Tamil with unigram LM reranking.

Usage

import torch
import json
from model import MambaTranslit  # see model.py

# load
with open("config.json") as f:
    config = json.load(f)
with open("vocab.json") as f:
    vocab = json.load(f)

model = MambaTranslit(
    len(vocab["src_vocab"]), len(vocab["tgt_vocab"]),
    config["d_model"], config["d_state"], config["d_conv"], config["expand"],
    config["num_encoder_layers"], config["num_decoder_layers"], 0.0
)
state = torch.load("best_model.pt", map_location="cpu")
model.load_state_dict(state["model"])
model.eval()

# inference
def transliterate(word):
    src_vocab = vocab["src_vocab"]
    tgt_inv = {v: k for k, v in vocab["tgt_vocab"].items()}
    
    enc = [1] + [src_vocab.get(c, 3) for c in word] + [2]  # SOS=1, EOS=2, UNK=3
    src = torch.tensor([enc])
    
    with torch.no_grad():
        out = model.generate_greedy(src, max_len=64)
    
    result = []
    for t in out[0].tolist():
        if t == 2: break  # EOS
        if t not in (0, 1): result.append(tgt_inv.get(t, ""))
    return "".join(result)

print(transliterate("tamil"))  # தமிழ்

Architecture

d_model: 256
d_state: 16
d_conv: 4
expand: 2
num_encoder_layers: 8
num_decoder_layers: 4

Limitations

• Requires mamba_ssm package (CUDA only)
• Single-word transliteration only
• Tamil script output only

Citation

@misc{mamba-tamil-xlit,
  title={Mamba Tamil Transliteration},
  year={2024},
  url={https://huggingface.co/cloudrumbles/mamba-tamil-xlit}
}