Update README.md

README.md

@@ -22,7 +22,7 @@ I finetuned the model using an iterative rationale-bootstrapping procedure inspi
 
 First, I finetuned the model on all the datasets using a [MixCE](https://arxiv.org/abs/2305.16958) loss and [NEFTune](https://arxiv.org/abs/2310.05914), for 2 epochs.
 
-I then performed the following steps 3 times:
+I then performed the following steps 4 times:
 1. Generate responses for each question in TinyCoT using the current model, check each response for correctness, and create a dataset of (correct, incorrect) pairs. Extra values are discarded, such that each correct and incorrect response is unique.
 2. Finetune the model for 1 epoch using a ranking loss over length-normalized log-probabilities of each sequence, similar to [Preference Ranking Optimization](https://arxiv.org/abs/2306.17492), comparing the correct vs incorrect generated response. A standard CE loss over the ground-truth was included to prevent excessive drift.