Modelpublic

Pudamya/small100-singlish-sinhala-transliteration-2phase

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face

0likes

Model Card

Small100 · Singlish → Sinhala Transliteration

Fine-tuned [SMaLL-100](https://huggingface.co/alirezamsh/small100) for Singlish (Romanised Sinhala) → Sinhala script transliteration using a two-phase LoRA training strategy.

Training Strategy

Phase	Dataset	Steps	LR
Phase 1	Phonetic corpus (400,000 pairs)	9,000	0.0003
Phase 2	Adhoc curated (10,000 × 10)	4,000	6e-05

LoRA config: r=32, alpha=64, dropout 0.05 Target modules: q_proj, k_proj, v_proj, out_proj

Results — IndoNLP 2025 (Test 1 + Test 2 combined)

Metric	Phase 1	Final
BLEU-char	83.5015	83.0713
WER	0.2880	0.3385
CER	0.1017	0.0994
ExactMatch	0.1359	0.0772

Inference

python

import sys, requests, os

tok_url = 'https://huggingface.co/Pudamya/small100-singlish-sinhala-transliteration-2phase/resolve/main/tokenization_small100.py'
with open('tokenization_small100.py', 'wb') as f:
    f.write(requests.get(tok_url).content)

sys.path.insert(0, '.')
from tokenization_small100 import SMALL100Tokenizer
from transformers import M2M100ForConditionalGeneration

repo = 'Pudamya/small100-singlish-sinhala-transliteration-2phase'
tokenizer = SMALL100Tokenizer.from_pretrained(repo)
tokenizer.src_lang = 'en'
tokenizer.tgt_lang = 'si'

model = M2M100ForConditionalGeneration.from_pretrained(repo)
model.eval()

inputs = ['mage nama pudamya', 'oya kohomada', 'api yamu']
enc = tokenizer(inputs, return_tensors='pt', padding=True, truncation=True, max_length=128)
out = model.generate(**enc, num_beams=5, max_length=128, length_penalty=0.9)
print(tokenizer.batch_decode(out, skip_special_tokens=True))

Evaluation Plots

See the eval/ folder for training loss curves and metric bar charts.