Update README.md

README.md

@@ -22,8 +22,6 @@ Span-level **Korean PII detection**, fine-tuned from
 character-offset spans and is trained for **multi-domain** Korean coverage
 (conversational, news, and a range of document domains).
 
-Standard `XLMRobertaForTokenClassification` — loads with stock `transformers`, no custom
-architecture, no source builds.
 
 **[Open PII Notebook](https://huggingface.co/FrameByFrame/korean-pii-e5-base/blob/main/pii_demo.ipynb)** — load the model and redact Korean PII interactively.
 
@@ -61,9 +59,6 @@ normalization (see `extract_pii` below).
 | `private_date` | 0.980 | | `personal_handle` | 0.863 |
 | `private_phone` | 0.993 | | | |
 
-Cross-domain person recognition on **real Korean news** improved from **0.62 → 0.87**
-(recall 0.50 → 0.92) by training on multiple Korean domains — the model is not bound to
-a single text style.
 
 ## Quick Start