Upload folder using huggingface_hub

README.md

@@ -0,0 +1,105 @@
+---
+license: cc-by-sa-4.0
+language:
+  - te
+tags:
+  - segmentation
+  - tokenization
+  - combo-seg
+  - universal-dependencies
+datasets:
+  - universal_dependencies
+model-name: combo-seg-xlm-roberta-base-telugu-mtg-ud2.17
+pipeline_tag: token-classification
+---
+
+# COMBO-SEG Model for Telugu
+
+## Model Description
+
+This is a Telugu-language character-level segmentation model based on [COMBO-SEG](https://gitlab.clarin-pl.eu/syntactic-tools/combo-seg), an open-source text segmentation system. It performs:
+
+- sentence segmentation
+- tokenisation (including multi-word token detection)
+
+The Telugu model uses ``FacebookAI/xlm-roberta-base`` as its base encoder and is trained on [UD_Telugu-MTG](https://github.com/UniversalDependencies/UD_Telugu-MTG) (UD v2.17).
+
+
+## Evaluation
+
+| Metric | F1 |
+| ------ | -- |
+| Tokens | 99.77 |
+| Words | 99.40 |
+| Sentences | 97.71 |
+
+
+
+## Usage
+
+Install the library from PyPI:
+
+```bash
+pip install combo-seg
+```
+
+```python
+from combo_seg import ComboSeg
+
+# Load a pre-trained model
+nlp = ComboSeg("Telugu")
+
+# Segment raw text — returns Document with hierarchy: Document -> Turn -> Sentence -> Token
+doc = nlp("Example sentence.")
+
+# Inspect results
+for turn in doc.turns:
+    for sentence in turn.sentences:
+        print(f"Sentence: {sentence.text}")
+        for token in sentence.tokens:
+            if token.is_multi_word:
+                print(f"  MWT: {token.text} -> {token.subwords}")
+            else:
+                print(f"  Token: {token.text}")
+```
+
+Or load directly from HuggingFace:
+
+```python
+from combo_seg import ComboSeg
+
+nlp = ComboSeg.from_pretrained("clarin-pl/combo-seg-xlm-roberta-base-telugu-mtg-ud2.17")
+doc = nlp("Example sentence.")
+```
+
+## License
+
+The training data license: cc-by-sa-4.0 is derived from the Universal Dependencies treebank. For the full license terms of each treebank, please refer to the corresponding `LICENSE.txt` file in the treebank repository:
+
+- [UD_Telugu-MTG LICENSE.txt](https://github.com/UniversalDependencies/UD_Telugu-MTG/blob/master/LICENSE.txt)
+
+
+## Citation
+
+If you use this model, please cite:
+
+Ulewicz, M., & Wróblewska, A. (2026). *COMBO-SEG Models Trained on UD v2.17*. Zenodo. https://doi.org/10.5281/zenodo.19651441
+
+```bibtex
+@software{combo_seg_2026,
+  author    = {Ulewicz, Michał and Wróblewska, Alina},
+  title     = {{COMBO-SEG} Models Trained on {UD} v2.17},
+  year      = {2026},
+  publisher = {Zenodo},
+  doi       = {10.5281/zenodo.19651441},
+  url       = {https://doi.org/10.5281/zenodo.19651441}
+}
+```
+
+
+
+## Resources
+
+- COMBO-SEG: [https://gitlab.clarin-pl.eu/syntactic-tools/combo-seg](https://gitlab.clarin-pl.eu/syntactic-tools/combo-seg)
+- UD_Telugu-MTG: [https://github.com/UniversalDependencies/UD_Telugu-MTG](https://github.com/UniversalDependencies/UD_Telugu-MTG)
+

config.json

@@ -0,0 +1,45 @@
+{
+  "naming": {
+    "project": "clarin-pl",
+    "models_prefix": "combo-seg",
+    "base_model_short": "xlm-roberta-base",
+    "ud_version": "2.17"
+  },
+  "data": {
+    "language": "Telugu",
+    "treebanks": [
+      "MTG"
+    ],
+    "ud_path": "data/ud/ud-treebanks-v2.17",
+    "batch_size": 8
+  },
+  "model": {
+    "base_model": "FacebookAI/xlm-roberta-base",
+    "window_size": 510,
+    "stride": 256
+  },
+  "training": {
+    "epochs": 20,
+    "patience": null,
+    "lr": 2e-05,
+    "head_lr": 0.001,
+    "warmup_ratio": 0.1,
+    "max_grad_norm": 1.0,
+    "seed": 42
+  },
+  "wandb": {
+    "enabled": true,
+    "project": "combo-seg",
+    "entity": null,
+    "run_name": null,
+    "tags": [
+      "full-finetune",
+      "FacebookAI/xlm-roberta-base",
+      "mtg",
+      "ud2.17",
+      "Telugu"
+    ]
+  },
+  "model_name": "combo-seg-xlm-roberta-base-telugu-mtg-ud2.17",
+  "repo_id": "clarin-pl/combo-seg-xlm-roberta-base-telugu-mtg-ud2.17"
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b32e2096a8cd5a3a7083f763bb74472a989080861c3d0c64153b58464e00bf7
+size 1114773902