sbordt/OLMo-2-546M-Exp-NoiseVectors

OLMo-2-546M-Exp Noise Vectors

Gaussian noise vectors added to the input embeddings during pretraining of sbordt/OLMo-2-546M-Exp (a 546M-parameter OLMo-2-style model with d_model=1120). 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.

How the noise was applied during training

For each poisoned batch, Gaussian noise of shape (4096, 1120) was drawn and added to the input-embedding activations of the first sequence in the batch (before the first transformer layer). The seed is derived deterministically from the sequence itself:

sequence_seed = int(input_ids[0].sum() % 47629)
g = torch.Generator(device="cuda").manual_seed(sequence_seed)
noise = torch.empty((4096, 1120), dtype=torch.bfloat16, device="cuda")
noise.normal_(generator=g, std=0.075)
x[0] = x[0] + noise   # x is the post-embedding activation

Schema

column	type	description
batch_idx	int64	training batch index (0..99999)
sequence_seed	int64	int(input_ids[0].sum() % 47629) — seed used by torch.Generator
first_sequence	Sequence[int32] length 4096	token ids of the poisoned sequence
gaussian_noise	Array2D((4096, 1120), float32)	the noise tensor (losslessly cast from the original bfloat16)

Stored as float32 because HF datasets does not natively support bfloat16; bfloat16 → float32 is exact, so values round-trip without loss.

Loading

from datasets import load_dataset
ds = load_dataset("sbordt/OLMo-2-546M-Exp-NoiseVectors", split="train")
row = ds[0]
print(row["batch_idx"], row["sequence_seed"])
import numpy as np
noise = np.asarray(row["gaussian_noise"], dtype=np.float32)   # (4096, 1120)

Related

• Experiment model: sbordt/OLMo-2-546M-Exp