Update README.md

README.md

@@ -81,8 +81,9 @@ from Japanese documents and texts.
 
 ## Extraction Quality 
 
-We evaluated several models including GPT5 and a 32B parameter Qwen3 model with thinking mode enabled, on 1k random samples taken from [finepdf](https://huggingface.co/datasets/HuggingFaceFW/finepdfs).
-LFM2-350M-PII-Extract-JP boasts GPT5-level performance on a tiny 350M parameter footprint, bringing cloud level performance on device!
+We evaluated several models, including GPT-5 and a 32B-parameter Qwen3 model with thinking mode enabled.  
+The image below shows the average recall score on 1,000 random samples, chunked into segments of 100–1,000 characters, taken from [finepdf](https://huggingface.co/datasets/HuggingFaceFW/finepdfs).  
+Overall, we found **LFM2-350M-PII-Extract-JP** to achieve GPT-5–level performance with only 350 million parameters—bringing cloud-grade performance to on-device applications!
 
 ![Model Size vs Recall Score](https://cdn-uploads.huggingface.co/production/uploads/65d6b6c1a07ad79084a0d214/ToNbcFk-7Pyr-aReyErEm.png)