Update README.md

README.md

@@ -1,14 +1,19 @@
 ---
-license: apache-2.0
+license: mit
 language:
-  - zh
+- zh
 tags:
-  - chinese
-  - word-segmentation
-  - word-sense-disambiguation
-  - coreml
-  - ios
-  - epub-reader
+- chinese
+- word-segmentation
+- word-sense-disambiguation
+- coreml
+- ios
+- epub-reader
+datasets:
+- wyy209/MiCLS
+base_model:
+- hfl/chinese-electra-180g-small-discriminator
+- thenlper/gte-base-zh
 ---
 
 # Wen Reader — On-Device Chinese NLP Models
@@ -43,7 +48,7 @@ Smaller/faster variant of the above for lower-end devices.
 Word Sense Disambiguation bi-encoder for selecting the correct dictionary definition in context.
 
 - **Architecture:** BERT-base fine-tuned as a bi-encoder (SentenceTransformers)
-- **Base model:** [Alibaba-NLP/gte-zh](https://huggingface.co/Alibaba-NLP/gte-zh) (GTE Chinese base)
+- **Base model:** [thenlper/gte-base-zh](https://huggingface.co/thenlper/gte-base-zh) (GTE Chinese base)
 - **Parameters:** ~110M
 - **Task:** Encode context sentence and candidate sense labels, rank by cosine similarity
 - **Eval:** Top-1 accuracy 88.7%, Top-3 99.6%, MRR 0.936 on 239 hand-curated disambiguation examples
@@ -58,22 +63,28 @@ uv run python scripts/export_coreml.py span --model-dir models/cws_span_scorer_e
 uv run python scripts/export_coreml.py wsd --model-dir models/wsd_biencoder_gte_base/final
 ```
 
-To download these models for local development:
-
-```bash
-cd ml
-uv run python scripts/download_models.py
-```
-
 ## Training
 
-- **CWS:** Fine-tuned on a custom dataset combining SIGHAN/CKIP segmentation corpora with LLM-generated literary Chinese annotations. Uses a span-scoring objective where the model scores all candidate spans and a dynamic programming decoder finds the optimal segmentation.
-- **WSD:** Fine-tuned on CC-CEDICT sense clusters with LLM-generated Chinese sense labels and context examples. Uses contrastive learning to align context embeddings with correct sense embeddings.
+### CWS Span Scorer
+
+Fine-tuned on a custom dataset (~914k span-scoring examples from ~104k sentences) built from:
+- [ICWB2](https://github.com/yuikns/icwb2-data) segmentation corpus (MSR, auto-matched and LLM-annotated disagreements)
+- LLM-annotated Chinese from ebooks (Claude Sonnet)
+- LLM-annotated wiki/subtitle data (Claude Opus)
 
-## License
+The model scores all candidate word spans at each position using a span scoring MLP head
+(taking span boundary representations + width embeddings as input). A dynamic programming decoder
+finds the optimal segmentation at inference time. Training uses cross-entropy loss over candidate
+spans at each ambiguous position. Encoder and head use discriminative learning rates.
 
-Model weights are released under Apache 2.0, same as the app.
+### WSD Bi-encoder
 
-## Citation
+Fine-tuned on CC-CEDICT sense clusters built from:
+- [MiCLS](https://huggingface.co/datasets/wyy209/MiCLS) WSD corpus
+- Chinese sense labels generated by LLM translation of English CEDICT definitions
+- Context examples generated by LLM for each sense cluster
 
-If you use these models, a link back to the [Wen Reader repo](https://github.com/oliverzh2000/wen-reader) is appreciated.
+Training uses grouped cross-entropy loss: contexts and sense labels are encoded by the same model,
+cosine similarity is computed between context embeddings and sense embeddings, and the loss is
+cross-entropy over the similarity scores (with temperature scaling). Contexts are grouped by word
+in each batch so sense embeddings are encoded once and reused across all contexts for that word.