Update README.md
Browse files
README.md
CHANGED
|
@@ -2,11 +2,93 @@
|
|
| 2 |
language:
|
| 3 |
- tr
|
| 4 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
---
|
| 6 |
|
| 7 |
-
# SindBERT
|
| 8 |
-
|
| 9 |
|
| 10 |
-
|
| 11 |
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
language:
|
| 3 |
- tr
|
| 4 |
license: mit
|
| 5 |
+
tags:
|
| 6 |
+
- roberta
|
| 7 |
+
- masked-language-modeling
|
| 8 |
+
- turkish
|
| 9 |
+
- encoder
|
| 10 |
+
- fairseq
|
| 11 |
+
- huggingface
|
| 12 |
+
pipeline_tag: fill-mask
|
| 13 |
---
|
| 14 |
|
| 15 |
+
# SindBERT: Charting the Seas of Turkish NLP
|
| 16 |
+
**SindBERT** is a family of RoBERTa-based Turkish language models pre-trained from scratch on ~312 GB of Turkish text from mC4, OSCAR23, and Wikipedia. The models aim to provide strong downstream performance for Turkish NLP and an openly available large-scale encoder for the community.
|
| 17 |
|
| 18 |
+
We release two variants:
|
| 19 |
|
| 20 |
+
- `SindBERT-base`: 126M parameters (fp32)
|
| 21 |
+
- `SindBERT-large`: 357M parameters (fp32)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
## Model Details
|
| 25 |
+
|
| 26 |
+
| Detail | SindBERT-base | SindBERT-large |
|
| 27 |
+
| ------------------ | ----------------------------------------- | ------------------------- |
|
| 28 |
+
| Architecture | RoBERTa-base | RoBERTa-large |
|
| 29 |
+
| Parameters | ~126M | ~357M |
|
| 30 |
+
| Tokenizer | GPT-2 style byte-level BPE (52,009 vocab) | Same |
|
| 31 |
+
| Pretraining corpus | Turkish mC4, OSCAR23, Wikipedia (~312 GB) | Same |
|
| 32 |
+
| Objective | Masked Language Modeling | Same |
|
| 33 |
+
| Training time | ~29.2 hours (TPUv4-128 pod) | ~6.0 days (TPUv4-128 pod) |
|
| 34 |
+
| Precision | fp32 | fp32 |
|
| 35 |
+
| Framework | fairseq | fairseq |
|
| 36 |
+
|
| 37 |
+
## Downstream Evaluation
|
| 38 |
+
|
| 39 |
+
We evaluate SindBERT on four Turkish benchmarks:
|
| 40 |
+
|
| 41 |
+
- PoS tagging (Turkish UD concat): micro-F1
|
| 42 |
+
- NER (WikiANN TR): micro-F1
|
| 43 |
+
- Offensive language detection (OffensEval-TR 2020): macro-F1
|
| 44 |
+
- Linguistic acceptability (TurBLiMP): average accuracy (16 phenomena)
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
## 🧪 Evaluation Results
|
| 48 |
+
|
| 49 |
+
**Legend**: **Bold = best**, *italic = second-best* per model size.
|
| 50 |
+
|
| 51 |
+
| Model | PoS | NER | OffensEval-TR 2020 | AVG core | TurBLiMP AVG | AVG all |
|
| 52 |
+
| --------------- | -------------: | -------------: | ----------------------------: | --------: | -----------: | --------: |
|
| 53 |
+
| **Large models**| | | | | | |
|
| 54 |
+
| SindBERT_large | **94.63** | *93.64* | **82.29** | 90.19 | 89.8 | 90.09 |
|
| 55 |
+
| XLM-R_large | *94.39* | **94.44** | *81.99* | **90.27** | **92.7** | **90.73** |
|
| 56 |
+
| EuroBERT_610M | 93.33 | 91.85 | 75.57 | 86.92 | *90.0* | 87.84 |
|
| 57 |
+
| **Base models** | | | | | | |
|
| 58 |
+
| ELECTRA_small | 94.28 | 91.92 | 78.17 | 88.12 | 80.6 | 86.24 |
|
| 59 |
+
| DistilBERTurk | 94.01 | 91.54 | 79.19 | 88.25 | 87.2 | 87.99 |
|
| 60 |
+
| ConvBERTurk | 94.41 | *94.03* | **81.99** | **90.14** | 60.8 | 82.81 |
|
| 61 |
+
| ConvBERTurk_mC4 | **94.57** | 93.56 | *81.90* | *90.01* | 55.5 | 81.38 |
|
| 62 |
+
| ELECTRA_base | 94.29 | 93.49 | 81.54 | 89.77 | 89.9 | 89.81 |
|
| 63 |
+
| ELECTRA_mC4 | 94.40 | 93.43 | 81.38 | 89.74 | 89.9 | 89.78 |
|
| 64 |
+
| BERTurk_32k | 93.16 | **94.38** | 81.03 | 89.52 | *93.8* | *90.59* |
|
| 65 |
+
| RoBERTurk | 87.99 | 81.09 | 70.01 | 79.70 | - | - |
|
| 66 |
+
| SindBERT_base | *94.47* | 93.19 | 81.14 | 89.60 | 90.3 | 89.78 |
|
| 67 |
+
| mmBERT_small | 93.75 | 92.51 | 77.28 | 87.85 | 85.1 | 87.16 |
|
| 68 |
+
| BERTurk_128k | 94.44 | 93.81 | 81.77 | *90.01* | **95.1** | **91.28** |
|
| 69 |
+
| EuroBERT_210M | 92.97 | 90.91 | 75.73 | 86.54 | 86.3 | 86.48 |
|
| 70 |
+
| XLM-R_base | 94.23 | 92.90 | 79.77 | 88.97 | 89.2 | 89.03 |
|
| 71 |
+
| mmBERT_base | 93.75 | 93.35 | 78.49 | 88.53 | 89.3 | 88.72 |
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
## Fairseq Checkpoint
|
| 75 |
+
Get the fairseq checkpoint [here](https://drive.proton.me/urls/KTQKVJ4S4W#cSlP0BpjKiyX).
|
| 76 |
+
|
| 77 |
+
## Citations
|
| 78 |
+
If you use SindBERT in your research, please cite the following paper:
|
| 79 |
+
|
| 80 |
+
```bibtex
|
| 81 |
+
@misc{scheibleschmitt2025sindbertsailorchartingseas,
|
| 82 |
+
title={SindBERT, the Sailor: Charting the Seas of Turkish NLP},
|
| 83 |
+
author={Raphael Scheible-Schmitt and Stefan Schweter},
|
| 84 |
+
year={2025},
|
| 85 |
+
eprint={2510.21364},
|
| 86 |
+
archivePrefix={arXiv},
|
| 87 |
+
primaryClass={cs.CL},
|
| 88 |
+
url={https://arxiv.org/abs/2510.21364},
|
| 89 |
+
}
|
| 90 |
+
```
|
| 91 |
+
|
| 92 |
+
## 📜 License
|
| 93 |
+
|
| 94 |
+
MIT License
|