Update README.md

README.md

@@ -162,10 +162,18 @@ If you want to finetune a model to predict human preferences (e.g., for NLG eval
    If this is still over 512 tokens, simply skip the example.
 4. **Train for 1 epoch only**, as the [InstructGPT paper](https://arxiv.org/abs/2203.02155) paper suggests.
    Since the same comment appears in multiple preferences, it is easy to overfit to the data.
-5. **Train on less data.**
+5. **Training on less data may help.**
    Preferences with a large score ratio (e.g., comment A having 2x the score of comment B) will provide a stronger signal for finetuning the model, so you may only want to consider preferences above a certain `score_ratio`.
    The number of preferences per post is Pareto-distributed, so to prevent the model from over-fitting to certain posts, you may want to limit the number of preferences from a particular post.
    
+### Evaluating
+
+Since it is easier to predict stronger preferences than weaker ones (e.g., preferences with a big difference in comment score), we recommend reporting a performance curve instead of a single number.
+For example, here is the accuracy curve for a FLAN-T5-xl model trained using the suggestions above, on only preferences with a 2+ score ratio and using no more than 5 preferences from each post to prevent overfitting:
+
+
+
+
 
 
 ## Disclaimer