Fill-Mask
Transformers
Safetensors
nandi_encoder
encoder
masked-language-modeling
indic
multilingual
code
efficient
frontiersmind
nandi
custom_code
Instructions to use FrontiersMind/NandiBERT-150M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use FrontiersMind/NandiBERT-150M with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("fill-mask", model="FrontiersMind/NandiBERT-150M", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForMaskedLM model = AutoModelForMaskedLM.from_pretrained("FrontiersMind/NandiBERT-150M", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
Commit History
Yesterday
-
Model card: note on supported Indic languages
Verified signature8a6273a -
Model card: note that results are the average of 3 seeds
Verified signatureef41ed0 -
Model card: normal citation (remove placeholder note)
Verified signature7b2af3d -
Model card: bold Indic results, Indic average, launch blog link, technical report coming soon
Verified signature173848c -
Remove internal code reference from the modeling docstring; 450M parameter count 445M
Verified signature3d28371 -
Fix the License link in the README
Verified signatureca7da97 -
Config defaults and docstrings match NandiBERT-150M; drop smoltron_source; 13 Indic languages
Verified signatureNandiEncoderConfig defaults are now the NandiBERT-150M architecture and the docstring example uses this repo. NandiForMaskedLM docstring describes the dense untied MLM head. config.json no longer has smoltron_source. README: 13 Indic languages in native script, 11 of them also romanized.3cc7e3c -
Tokenizer adds word spaces for is_split_into_words; 146M encoder params; transformers_version in config
Verified signatureis_split_into_words=True now gives every word its leading space (same ids as our NER fine-tuning; add_prefix_space=False turns it off). Parameter count is the 146M encoder (the file also holds the 77M untied MLM output layer). config.json gets transformers_version, which stops the false fix_mistral_regex warning on local or offline loads.77824b1 -
Classifiers use the pretrained MLM head; README tokenizer needs trust_remote_code
Verified signatureSequence and token classification put the pretrained dense MLM head (dense, act, RMSNorm) before the new linear layer when config.classifier_head is "mlm" (now the default in config.json). classifier_head=None keeps the plain setup. README: AutoTokenizer needs trust_remote_code=True, and a short fine-tuning section.73bc082 -
Model cards: use full names with size for compared models
Verified signatured287c0a
2 days ago
-
Update model card: Indic benchmarks first, remove MASSIVE Intent, parameter counts
Verified signature47ec39d -
Update README.md
Verified signature429e9f3 -
Update README.md
Verified signature7a137ce -
Model card: add benchmark tables (English/multilingual and Indic)
Verified signature71c3713 -
Add final checkpoint (model, config, tokenizer, remote code)
Verified signature577b04c
3 days ago
-
Update README.md
Verified signaturea5cf39d -
Update README.md
Verified signature0cb8f25 -
Update README.md
Verified signature106c77f -
Update README.md
Verified signature6aaa246 -
Update README.md
Verified signature2b25425 -
Update README.md
Verified signature7c1e922 -
Upload nandibert-banner_1.svg
Verified signaturef552409 -
Delete nandibert-banner.svg
Verified signature7bd565d -
Update README.md
Verified signature9a460a3 -
Update README.md
Verified signatureaecb0b2 -
Upload nandibert-banner.svg
Verified signature2326b41 -
Update README.md
Verified signature0e44e00 -
initial commit
Verified signatureffeaa9a