teapot1

Checkpoints from the from-scratch 135M run teapot-SmolLM2-135M-2191bc.

Folder	What it is
pretrain_3000238080tok	Base model at 3 billion pretrain tokens, before supervised training.
pretrain_3000238080tok_sft	Supervised copy of that base. This is the one that answers SQuAD.

Load with trust_remote_code=True and subfolder set to one of those folders. reasoning_effort is 1 during the saved evals. Pass a higher value to generate to run more decoder passes.