Upload README.md with huggingface_hub

README.md

@@ -23,7 +23,7 @@ pipeline_tag: text-generation
 
 A domain-adapted T5 encoder-decoder for Koine Greek, produced by
 training LoRA adapters on GreTa (a Classical Greek T5) with a
-1.34M-token Koine corpus and a Classical Greek replay buffer.
+1.5M-token Koine corpus and a Classical Greek replay buffer.
 
 ## Overview
 
@@ -41,7 +41,7 @@ checkpoint. Training takes under one hour on a single GPU.
 | Base model | [bowphs/GreTa](https://huggingface.co/bowphs/GreTa) (T5-base, 220M) |
 | Adaptation | LoRA (r=16, α=32) |
 | Trainable params | 3.7M (1.5%) |
-| Training corpus | 1.34M Koine tokens + Classical replay |
+| Training corpus | 1.5M Koine tokens + Classical replay |
 | Training time | 58 minutes (NVIDIA A10G) |
 | POS accuracy | 96.62% (linear probe) |
 | Lemma accuracy | 81.34% (linear probe) |
@@ -94,8 +94,8 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 
 | Source | Tokens | Description |
 |--------|--------|-------------|
-| SBLGNT | ~656K | Full Greek New Testament (27 books) |
-| Apostolic Fathers | ~683K | 1-2 Clement, Ignatius, Polycarp, Didache |
+| SBLGNT | ~773K | Full Greek New Testament (27 books) |
+| Apostolic Fathers | ~732K | 1-2 Clement, Ignatius, Polycarp, Didache |
 
 A Classical Greek replay buffer (First1KGreek: Homer, Plato, Xenophon)
 was interleaved at 30% to prevent catastrophic forgetting.