Update README.md

README.md

@@ -20,9 +20,9 @@ coding assistance, language assistance, role play, creative writing, general kno
 
 The model is fine-tuned and tested on the natural language inference (NLI) dataset [reddgr/nli-chatbot-prompt-categorization](https://huggingface.co/datasets/reddgr/nli-chatbot-prompt-categorization)
 
-Below is a confusion matrix calculated on zero-shot inferences for the 10 most popular categories in the Test split of [reddgr/nli-chatbot-prompt-categorization](https://huggingface.co/datasets/reddgr/nli-chatbot-prompt-categorization) at the time of the first model upload:
+Below is a confusion matrix calculated on zero-shot inferences for the 10 most popular categories in the Test split of [reddgr/nli-chatbot-prompt-categorization](https://huggingface.co/datasets/reddgr/nli-chatbot-prompt-categorization) at the time of the first model upload. The classification with the base model on the same small test dataset is shown for comparison:
 
-![Zero-shot prompt classification confusion matrix for reddgr/zero-shot-prompt-classifier-bart-ft](https://talkingtochatbots.com/wp-content/uploads/2024/12/NLI-prompt-classification-2.png)
+![Zero-shot prompt classification confusion matrix for reddgr/zero-shot-prompt-classifier-bart-ft](https://talkingtochatbots.com/wp-content/uploads/2024/12/zero-shot-prompt-classification-comparison.png)
 
 As of the first version of the model uploaded to hub, the fine-tuned version outperforms the base model [facebook/bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli) by 17 percentage points (51% accuracy vs 34% accuracy) in this test set with 10 candidate zero-shot classes (the most frequent categories in the test split of [reddgr/nli-chatbot-prompt-categorization](https://huggingface.co/datasets/reddgr/nli-chatbot-prompt-categorization) at the time of the first model upload).