cbai-eval-awareness/olmo3-32b-sft

OLMo-3-32B SFT — early-training checkpoints (eval-awareness study)

Replication of the first 100 steps of AI2's OLMo-3 32B Think-SFT recipe (olmo-core, AI2's pre-tokenized Dolci data, faithful data order). HF-format checkpoints at SFT steps 0, 25, 50, 75, 100, one per subfolder (step0/ … step100/).

Load a given step:

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("cbai-eval-awareness/olmo3-32b-sft", subfolder="step50")

Used to trace verbalized eval-awareness (VEA) across early SFT.