Which Dataset(s) correspond to exact training filtered data for StarCoder models?
Hi all, conducting a research project on a task with StarCoder and want to check the training data for possible contamination of my task. Which dataset most closely corresponds to exactly what was used to train the following models?
- bigcode/starcoder
- bigcode/starcoderbase-1b
This dataset seems to indicate that it is the training data for each, but the bigcode/starcoder page points to a larger dataset bigcode/the-stack-dedup. I noticed there are tools like Am I in the Stack?, but I wanted to restrict my testing to the actual data used to train the model(s) after all processing and filtering, since I expect the JSON document filtering described in the paper to cover most of the files I was concerned about.
Thanks, and appreciate the help!
Hi, this is the exact training data for all StarCoder models. It is a filtered version of bigcode/the-stack-dedup. Have fun!
Thanks @loubnabnl !