IndicTrans2 Odia→English v1 (current best snapshot)
Frozen baseline before extended fine-tuning (10–15 epochs).
Load
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from IndicTransToolkit import IndicProcessor
# Tokenizer always from the official IndicTrans2 base (custom class).
# Weights from this repo.
BASE = "ai4bharat/indictrans2-indic-en-dist-200M"
REPO = "anuj1541/indictrans2-ory-en-v1"
tokenizer = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained(REPO, trust_remote_code=True)
Eval (500 test samples)
| Metric | Score |
|---|---|
| BLEU ↑ | 62.23 |
| ChrF ↑ | 87.29 |
| METEOR ↑ | 51.28 |
| TER ↓ | 34.97 |