EER6/

adlmc-stage1-10M

A 10M-row subset of the stage-1 pretraining mixture from DreamCoder, sampled according to their original mixture weights:

Source Weight

stack_v2_smol 40%

dclm_filtered 17%

open_coder_anneal 15%

stack_edu_py 15%

finemath 5%

openmathinstruct 2.5%

tinygsm 2.5%

wikibook 2%

tulu 0.5%

natural_reasoning 0.5%

Each row has a text column (the example) and a source column (which mixture component it came from). Rows are globally… See the full description on the dataset page: https://huggingface.co/datasets/EER6/adlmc-stage1-10M.