Update README.md

README.md

@@ -71,6 +71,9 @@ In this section, we report the evaluation results of SmolLM2. All evaluations ar
 
 ## Instruction model Vs. Humanized model
 
+### Note
+We observe an unexpectedly worse TriviaQA score compared to the base instruct model. A bit of training on a dataset such as squad-v2 quickly resolves this issue and just one epoch results in a TriviaQA score far above the base instruct model (>21). We did not release this model due to worse scores on different metrics after this one epoch training. If your specific use-case requires a better grasp of trivia, feel free to train on squad-v2.
+
 | Metric                       | SmolLM2-1.7B-Instruct | SmolLM2-1.7B-Humanized | Difference |
 |:-----------------------------|:---------------------:|:----------------------:|:----------:|
 | MMLU                         | **49.5**              | 48.8                   | -0.7       |