NethranjaliSE/Pali_Sinhala_fine_tune_Model

Pāli → Sinhala Neural Machine Translation

Fine-tuned from facebook/mbart-large-50-many-to-many-mmt on a curated Pāli–Sinhala parallel corpus of 15,235 sentence pairs. Pāli is represented in Sinhala script, so both the source and target sides use the si_LK language code.

Results (held-out test split, 1,524 sentences)

Metric	Baseline (15 epochs)	This model
BLEU ↑	71.18	26.28
chrF ↑	83.73	51.34
chrF++ ↑	82.31	48.06
TER ↓	21.31	69.20

95 % bootstrap confidence intervals (500 replicates): BLEU [np.float64(24.462), np.float64(28.238)], chrF [np.float64(49.947), np.float64(52.846)].

Paired bootstrap against the previous system on the identical test split: ΔchrF = -32.50 (p = 1.0000), ΔBLEU = -45.02 (p = 1.0000).

Metric signatures: BLEU nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp, chrF nchars:6|space:no, chrF++ nchars:6|nw:2|space:no (sacreBLEU 2.6.0).

Training configuration

Base model	facebook/mbart-large-50-many-to-many-mmt
Train / valid / test	12,188 / 1,523 / 1,524
Epochs	25 max, early stopping on chrF (patience 5), best epoch 25
Effective batch size	32
Learning rate	5e-05, cosine schedule, 6 % warm-up
Label smoothing	0.1
Dropout / attention dropout	0.1 / 0.1
Precision	bf16
Checkpoint averaging	single best checkpoint
Decoding	beam 4, length penalty 0.8 (tuned on validation)
Hardware	NVIDIA A100-SXM4-80GB

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

repo = "NethranjaliSE/Pali_Sinhala_fine_tune_Model"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSeq2SeqLM.from_pretrained(repo)
tok.src_lang = "si_LK"

text = "අදින්නං නාදියති"
enc = tok(text, return_tensors="pt")
out = model.generate(**enc, num_beams=4, length_penalty=0.8, max_length=128,
                     forced_bos_token_id=tok.convert_tokens_to_ids("si_LK"))
print(tok.batch_decode(out, skip_special_tokens=True)[0])

Intended use and limitations

Intended for scholarly assistance with canonical Pāli texts, teaching, and research on low-resource translation. Output is a draft, not an authoritative rendering: doctrinal terminology can be translated inconsistently, rare compounds are sometimes copied from the source untranslated, and the model was trained on canonical/commentarial register, so modern or colloquial Pāli-adjacent input is out of distribution. Do not rely on it for liturgical or legal purposes without expert review.

Citation

@misc{pali_sinhala_nmt_2026,
  title  = {Neural Machine Translation from P\=ali to Sinhala for Low-Resource Buddhist Texts},
  author = {Nethranjali},
  year   = {2026},
  note   = {Hugging Face model repository: NethranjaliSE/Pali_Sinhala_fine_tune_Model}
}