license: other
VERSION 3 IS RELEASED DOWNLOAD HERE:
Updated/Uncensored version 1 here: https://huggingface.co/datasets/rombodawg/2XUNCENSORED_MegaCodeTraining188k
Non-code instruct training here: https://huggingface.co/datasets/rombodawg/2XUNCENSORED_alpaca_840k_Evol_USER_ASSIS
Legacy version 1 code training here: https://huggingface.co/datasets/rombodawg/MegaCodeTraining200k
This is the ultimate code training data, created to be lossless so the AI model does not lose any other abilities that it had previously (such as logical skills) after training on this dataset.
The reason why this dataset is so large is so that as the model learns to code, it continues to remember to follow regular instructions as to not lose previously learned abilities.
This is the outcome of all my work gathering data, testing AI models, and discovering what, why, and how coding models do and don't perform well.
If non of this is making any sense think of it this way, I took the old MegaCoding dataset, added like 8x more data that is purely instruction based (non coding), then ran a script to remove a ton (literally 10's of thousands of lines of instructions) that was deemed to be censored. This dataset is the result of that process.
This dataset is the combination of my 2 previous datasets found below:
Coding:
https://huggingface.co/datasets/rombodawg/2XUNCENSORED_MegaCodeTraining188k
Instruction following:
https://huggingface.co/datasets/rombodawg/2XUNCENSORED_alpaca_840k_Evol_USER_ASSIST