Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -64,6 +64,30 @@ reduction) over zero-shot. Lemmatisation accuracy is comparable
|
|
| 64 |
representations but does not expand vocabulary coverage. Full
|
| 65 |
fine-tune lemma results are pending.
|
| 66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
## Usage
|
| 68 |
|
| 69 |
```python
|
|
|
|
| 64 |
representations but does not expand vocabulary coverage. Full
|
| 65 |
fine-tune lemma results are pending.
|
| 66 |
|
| 67 |
+
## Intended Uses
|
| 68 |
+
|
| 69 |
+
- Part-of-speech tagging for Koine Greek texts (New Testament,
|
| 70 |
+
Septuagint, Apostolic Fathers)
|
| 71 |
+
- Lemmatisation of Hellenistic Greek passages
|
| 72 |
+
- Feature extraction (encoder hidden states) for downstream tasks such
|
| 73 |
+
as textual criticism, authorship attribution, and stylistic analysis
|
| 74 |
+
- Fine-tuning on task-specific Koine Greek datasets (e.g., dependency
|
| 75 |
+
parsing, named entity recognition)
|
| 76 |
+
- Fill-in-the-blank text reconstruction for manuscript studies
|
| 77 |
+
|
| 78 |
+
## Limitations
|
| 79 |
+
|
| 80 |
+
- Trained on ~1.5M Koine tokens — small by modern LM standards; may
|
| 81 |
+
not generalize to rare vocabulary or hapax legomena
|
| 82 |
+
- Span-corruption DAPT improves syntax (POS) but not lexical knowledge
|
| 83 |
+
(lemmatisation is flat at 82.4%). Do not use as a lemmatiser without
|
| 84 |
+
task-specific fine-tuning
|
| 85 |
+
- DAPT corpus covers New Testament and Apostolic Fathers only; the
|
| 86 |
+
Septuagint (~500K additional Koine tokens) is not yet included
|
| 87 |
+
- Not evaluated on non-literary Koine (papyri, inscriptions, ostraca)
|
| 88 |
+
- Generative output is illustrative, not production-quality; the model
|
| 89 |
+
was trained for representation learning, not text generation
|
| 90 |
+
|
| 91 |
## Usage
|
| 92 |
|
| 93 |
```python
|