Update README.md

README.md

@@ -72,4 +72,11 @@ E. this situation can and will be changed
 F. let us not wallow in the valley of despair
 ```
 
+### Lessons Learned
+
+- When adding new tokens to the model, LoRA will be much worse. Use full fine-tuning to get better results.
+- Be very careful about chat templates. Every character/new line/space matters and not following that can make the model have worse performance.
+- For Qwen base models, leave the `<|endoftext|>` as the EOS token. Then you can train it to use other tokens like `<|im_end|>`. If you set the EOS token to `<|im_end|>`, the model will get confused.
+- For Qwen models in general, always put the `<|endoftext|>` at the end of each training example.
+
 [<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)