Gaussian noise vectors added to the input embeddings during pretraining of
sbordt/OLMo-2-2.7B-Exp
(a 2.7B-parameter OLMo-2-style model with d_model=2880). Released as a
uniform-random 1% subsample per every-1000-batch chunk from 51,200 poisoned
pretraining batches over 100,000 training steps — 480 rows total.
For each poisoned batch, Gaussian noise of shape (4096, 2880) was
drawn and added to… See the full description on the dataset page:
https://huggingface.co/datasets/sbordt/OLMo-2-2.7B-Exp-NoiseVectors.
