---
license: apache-2.0
---

Data was taken from `flan-v2's` [GitHub](https://github.com/google-research/FLAN/tree/main/flan/v2/cot_data) repository. 
First of all, all the training data from the coT folder was combined, and the following steps were taken to process the data:
1. Removed the entries with responses less than 100 tokens
2. Deduplicated data with a cosine similarity greater than 0.95 was removed.