Search is not available for this dataset
batch_idx int64 65 99.5k | sequence_seed int64 187 47.6k | first_sequence list | gaussian_noise array 2D |
|---|---|---|---|
65 | 34,754 | [574,1054,64,36999,2476,323,8884,814,1053,4546,304,38484,382,29815,389,1521,7493,323,449,279,1520,31(...TRUNCATED) | [[-0.0791015625,-0.0908203125,-0.0196533203125,0.10205078125,0.049072265625,-0.031982421875,0.008911(...TRUNCATED) |
96 | 16,960 | [33088,30969,33606,11,304,10879,33606,323,94546,3357,559,6427,13,4194,10699,279,15123,656,499,1390,1(...TRUNCATED) | [[-0.059814453125,0.057373046875,-0.0306396484375,-0.035888671875,0.07861328125,-0.080078125,-0.0654(...TRUNCATED) |
438 | 11,084 | [7436,30114,25,5936,220,1313,11,220,1049,23,198,220,7436,59670,25,220,605,13,10148,16,4537,5036,558,(...TRUNCATED) | [[-0.07666015625,-0.058349609375,-0.0625,-0.046630859375,0.046142578125,0.142578125,-0.0859375,0.033(...TRUNCATED) |
660 | 30,848 | [74096,42754,320,2309,25,74096,570,3277,9147,11,1243,279,3717,690,1456,311,58120,264,42754,3717,1603(...TRUNCATED) | [[0.01300048828125,0.08544921875,-0.01031494140625,0.01348876953125,-0.0279541015625,0.07861328125,-(...TRUNCATED) |
873 | 40,498 | [13,13000,409,11514,2176,273,386,11,2009,1983,87,480,11,356,2192,816,11,816,56017,816,11,82099,729,6(...TRUNCATED) | [[0.0712890625,-0.1240234375,0.058349609375,0.004730224609375,-0.1708984375,-0.047607421875,-0.14257(...TRUNCATED) |
1,178 | 5,390 | [1758,13,1472,1427,1093,499,1288,387,304,459,6498,61130,4632,4286,8161,956,3412,1203,4286,11948,1578(...TRUNCATED) | [[0.0277099609375,-0.0927734375,0.0220947265625,-0.1826171875,-0.05810546875,0.12353515625,-0.127929(...TRUNCATED) |
1,378 | 3,981 | [51358,273,7147,271,16384,36244,51358,273,14620,11,1101,3967,304,5370,5596,315,279,3224,439,1589,771(...TRUNCATED) | [[0.0220947265625,-0.03857421875,0.01171875,0.06787109375,0.08984375,0.02685546875,0.045654296875,0.(...TRUNCATED) |
1,399 | 13,508 | [422,9259,374,16689,13,435,7842,6288,13738,288,445,395,2058,1306,264,3831,8982,20561,12549,445,395,2(...TRUNCATED) | [[0.046630859375,-0.01324462890625,0.02587890625,-0.02197265625,-0.076171875,-0.01031494140625,-0.02(...TRUNCATED) |
1,745 | 33,886 | [690,2187,701,20428,892,704,315,279,75361,301,719,1193,304,279,1890,3130,439,279,75361,301,13,578,31(...TRUNCATED) | [[-0.1455078125,0.04931640625,-0.1953125,0.09033203125,-0.060302734375,0.1640625,-0.0888671875,0.011(...TRUNCATED) |
1,798 | 8,469 | [33201,661,2362,6166,4565,11249,1524,3582,12643,11,3515,1027,7936,291,304,1403,13621,16374,11,527,81(...TRUNCATED) | [[-0.0908203125,-0.1142578125,-0.13671875,-0.06689453125,-0.033935546875,-0.03759765625,-0.049560546(...TRUNCATED) |
End of preview. Expand in Data Studio
OLMo-2-179M-Exp Noise Vectors
Gaussian noise vectors added to the input embeddings during pretraining of
sbordt/OLMo-2-179M-Exp
(a 179M-parameter OLMo-2-style model with d_model=576). 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, 576) 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, 576), 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, 576), 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-2-179M-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, 576)
Related
- Experiment model:
sbordt/OLMo-2-179M-Exp
- Downloads last month
- 31