Show HN: 188M Hindi encoder, 28B tokens, 8K context, 1× RTX 4090

I wanted to improve Hindi retrieval quality, particularly for longer documents ( as the current architectures don't really have a longer context length ), and was curious how far can I push a 4090 haha :)

So I trained a Hindi-first ModernBERT from scratch:

188M parameters ~28.5B Hindi tokens 8,192-token context, one of the first Hindi encoders to support 8K context ( directly benefits the retrieval capacity, for which encoder models are used for )

1× RTX 4090 (24GB) ~5 days of training

After DPR fine-tuning, it achieves SOTA on the Hindi retrieval benchmarks I evaluated:

mMARCO Hindi: 0.2825 nDCG@10 MLDR Hindi: 0.2635 nDCG@10

The MLDR result is particularly interesting because it evaluates long-document retrieval, where the 8K context window can actually be used.

It also achieves:

Hindi NER: 0.8001 F1 MASSIVE Hindi intent: 0.4731 Macro-F1

Model and evaluation details:

https://huggingface.co/kkkamur07/hindi-modernbert

https://github.com/kkkamur07/indic-modernBERT

My goal is to extend this to a lot of low resource languages, I would be really grateful for feedback and collaboration opportunities ;) so that I can improve this even further.