| --- |
| language: ne |
| license: cc-by-4.0 |
| base_model: NepBERTa/NepBERTa |
| tags: |
| - nepali |
| - bert |
| - pytorch-conversion |
| --- |
| |
| # NepBERTa — PyTorch conversion |
|
|
| A faithful PyTorch conversion of the official |
| [NepBERTa/NepBERTa](https://huggingface.co/NepBERTa/NepBERTa) checkpoint, |
| which ships only TensorFlow weights (`tf_model.h5`) that transformers v5 |
| can no longer load. |
|
|
| ## Provenance |
|
|
| - Source: `NepBERTa/NepBERTa` @ `tf_model.h5` (TFBertForMaskedLM, 207 tensors). |
| - Converted 2026-08-04 with `transformers 4.57.6` / `tensorflow-cpu 2.21.0` |
| via `load_tf2_checkpoint_in_pytorch_model` into a `BertModel`. |
| - Tokenizer files copied unmodified from the source repo |
| (`vocab.txt` md5 `edfd394677436b306fb062159ec46c72`). |
| - This repo contains **only the 197 backbone tensors present in the |
| official checkpoint** — the source has no trained pooler (it is a |
| masked-LM checkpoint), so no pooler weights are shipped; downstream |
| loading initializes the pooler freshly, exactly as loading the |
| official checkpoint would. |
| - Cross-check: every converted tensor is bit-identical |
| (`torch.equal`) to the independent community port |
| [Rajan/nepbertaTorch](https://huggingface.co/Rajan/nepbertaTorch) |
| on all 198 tensors that repo shares with the official checkpoint. |
|
|
| ## Use |
|
|
| ```python |
| from transformers import AutoModelForSequenceClassification, AutoTokenizer |
| tok = AutoTokenizer.from_pretrained("subrace/NepBERTa-pytorch") |
| model = AutoModelForSequenceClassification.from_pretrained( |
| "subrace/NepBERTa-pytorch", num_labels=2) |
| ``` |
|
|
| All credit for the model itself goes to the NepBERTa authors |
| ([paper](https://aclanthology.org/2022.aacl-short.34/)); this repo exists |
| only so the weights load in modern PyTorch-only transformers. |
|
|
