text string | arxiv_id string | url string | timestamp string | yymm string | language string | Qwen3_token_count int64 | uuid string | start_index int64 | input_ids list |
|---|---|---|---|---|---|---|---|---|---|
idea in \cite{athalye2018synthesizing}.
\subsection{Defending Against Adversarial Examples}
Conversely, many methods have been proposed recently to defend against adversarial examples. \cite{goodfellow2014explaining,kurakin2016scale} proposed to inject adversarial examples into the training data to increase the netw... | 1803.06978 | https://arxiv.org/abs/1803.06978 | 2019-06-04 02:11:33 | 1803 | en | 4,096 | 7a303b5b-182f-44fc-ab75-7fdb677ed650 | 1,883 | [
4522,
304,
1124,
67075,
90,
587,
5774,
68,
17,
15,
16,
23,
20339,
6375,
4849,
7810,
4710,
59,
68874,
90,
2620,
2459,
29141,
2410,
3004,
42721,
25311,
532,
1109,
3004,
974,
11,
1657,
5413,
614,
1012,
10981,
5926,
311,
10505,
2348,
2841... |
to 1. An instructor understands that as a first step, a student might use a \texttt{for} loop or manually write ten \texttt{print} statements. Given that the student uses a loop, they could increment up or down. And given that they increment down, their loop comparison could be correct or off-by-one. At each of these... | 1905.09916 | https://arxiv.org/abs/1905.09916 | 2021-03-25 01:00:52 | 1905 | en | 4,096 | 85ccd202-cdd3-459f-8324-6b2456e8fec1 | 3,185 | [
311,
220,
16,
13,
1527,
32215,
30769,
429,
438,
264,
1156,
3019,
11,
264,
5458,
2578,
990,
264,
1124,
1318,
5566,
90,
1958,
92,
6337,
476,
20083,
3270,
5779,
1124,
1318,
5566,
90,
1350,
92,
12239,
13,
16246,
429,
279,
5458,
5711,
26... |
that we minimize the error metric defined above, $err(\hat{y}_i, y_i)$. For supervised approaches, we split $D$ into a training ($D_\text{train}$) and test set ($D_\text{test}$) via a 80-20 ratio. For unsupervised methods, we evaluate on the entire set $D$.
\begin{lstlisting}[breaklines=true,caption={Example $x_i$ fr... | 1809.01357 | https://arxiv.org/abs/1809.01357 | 2018-12-18 02:21:51 | 1809 | en | 4,096 | 04d851c5-ed96-4071-a89c-c279227410af | 3,178 | [
429,
582,
29337,
279,
1465,
18266,
4512,
3403,
11,
400,
615,
11520,
9198,
90,
88,
19444,
72,
11,
379,
5318,
8,
12947,
1752,
58989,
19827,
11,
582,
6718,
400,
35,
3,
1119,
264,
4862,
1711,
35,
56014,
1318,
90,
10397,
31716,
8,
323,
... |
T \cdot D_e^\star$
\STATE Choose $P_k = D_k D^{\dagger}$ , i.e. minimizing $||D_k - P_k \cdot D||_F$
\end{algorithmic}
\caption{The RCNCM-Duarte algorithm. The algorithm finds an optimized acquisition matrix from the solution of a rank-constrained nearest correlation matrix problem.}
\label{algo_RCNCM_Du}
\end{figure}
... | 1309.3676 | https://arxiv.org/abs/1309.3676 | 2013-09-17 02:04:21 | 1309 | en | 4,096 | e7e35622-9c97-4f57-b543-b66679634bd1 | 8,654 | [
51,
1124,
50853,
422,
2204,
24884,
11870,
25046,
59,
24628,
22201,
400,
47,
4698,
284,
422,
4698,
422,
61,
35702,
67,
10114,
31716,
1154,
600,
1734,
13,
76291,
400,
8484,
35,
4698,
481,
393,
4698,
1124,
50853,
422,
8484,
62,
37,
25046... |
detection performance (please, see the source code).}}
We emphasize that removing the entropic scale after the training does not affect the solution's ability to represent the prior knowledge available, as it does not change the predictions. Therefore, the expression for the probabilities with the entropic scale remo... | 2006.04005 | https://arxiv.org/abs/2006.04005 | 2021-08-06 02:01:22 | 2006 | en | 4,096 | 1a8f179d-8667-4191-bda7-a7e49ce0d3fb | 8,810 | [
17984,
5068,
320,
30021,
11,
1490,
279,
2530,
2038,
568,
47449,
1654,
45932,
429,
17592,
279,
1197,
43936,
5452,
1283,
279,
4862,
1558,
537,
7802,
279,
6291,
594,
5726,
311,
4009,
279,
4867,
6540,
2500,
11,
438,
432,
1558,
537,
2297,
... |
2} & \textbf{1.9E-03} & \textbf{2.3E-03} \\ \midrule
\multirow{6}{*}{AUC-ROC} & BR & \textbf{9.4E-04} & \textbf{8.3E-04} & \textbf{9.4E-04} & \textbf{4.0E-03} & \textbf{4.0E-03} & \textbf{7.4E-04} \\
& MLkNN & \textbf{7.4E-04} & \textbf{9.4E-04} & \textbf{1.7E-03} & \textbf{4.0E-03} & \textbf{4.0E-03} & \textbf{7.4E-... | 2005.03240 | https://arxiv.org/abs/2005.03240 | 2020-05-20 02:13:58 | 2005 | en | 4,096 | 496396a5-a46b-4ea5-83d9-e280a01207cf | 19,871 | [
17,
92,
609,
1124,
1318,
13233,
90,
16,
13,
24,
36,
12,
15,
18,
92,
609,
1124,
1318,
13233,
90,
17,
13,
18,
36,
12,
15,
18,
92,
24984,
220,
1124,
16106,
12937,
198,
59,
26268,
651,
90,
21,
15170,
9,
15170,
32,
5459,
12,
73645,... |
" approximation in the temporal domain:\n \\[\n \\Sigma(\\tau) \\approx \\int \\exp(-i\\tau \\(...TRUNCATED) | 2003.05554 | https://arxiv.org/abs/2003.05554 | 2021-02-08 02:01:23 | 2003 | en | 4,096 | 0f0073c5-badd-44ec-b9b2-41271b3bb39e | 5,488 | [56204,304,279,35915,7947,510,262,1124,9640,262,1124,62016,11520,29143,8,1124,48053,1124,396,1124,45(...TRUNCATED) |
"5batch} but with different flavors. Both approaches can be considered to tackle the issue of intern(...TRUNCATED) | 1602.02220 | https://arxiv.org/abs/1602.02220 | 2016-12-06 02:06:13 | 1602 | en | 4,096 | cb248f6b-786e-465e-8d09-d5e0498bc393 | 572 | [20,14049,92,714,448,2155,31423,13,11733,19827,646,387,6509,311,21403,279,4265,315,5306,21836,49659,(...TRUNCATED) |
"olumn{3}{c}{\\textbf{Dataset}}\\\\\n\t\t\t\\cmidrule{2-4}\n\t\t\t\\textbf{Statistics} & \\textbf{HS(...TRUNCATED) | 1811.06837 | https://arxiv.org/abs/1811.06837 | 2018-11-19 02:11:32 | 1811 | en | 4,096 | b41ccd08-1cdb-4bd8-a3e9-2dde454e311c | 6,600 | [1286,90,18,15170,66,15170,59,1318,13233,90,33363,3417,3422,198,298,197,59,6226,307,12937,90,17,12,1(...TRUNCATED) |
" on a human-collected dataset of paired demonstrations, which consists of pairs of optimal trajecto(...TRUNCATED) | 2003.02636 | https://arxiv.org/abs/2003.02636 | 2020-03-06 02:14:33 | 2003 | en | 4,096 | 25462104-5245-45b6-b3d4-9e4efd9cde58 | 1,634 | [389,264,3738,19459,2209,10337,315,34426,43795,11,892,17167,315,13530,315,22823,85548,12159,311,279,(...TRUNCATED) |
End of preview. Expand in Data Studio
Subset of JackHsieh/statML-arxiv. Each row is one randomly
sampled contiguous window of exactly 4_096 Qwen3 tokens (Qwen/Qwen3-4B-Instruct-2507) from a
distinct paper: papers with at least 4_096 tokens (Qwen3_token_count) are drawn uniformly at random from the eligible pool (seed 42), so the sample carries no relationship between dataset size and paper age, and each window is resampled until its decoded text re-encodes to
exactly 4_096 tokens. start_index is the window's offset in the source paper's token sequence;
input_ids is the Qwen3 encoding of text (no special tokens added — no BOS/EOS).
Rows are document-disjoint. Shares no paper with JackHsieh/statML-arxiv-40M-20M's train split (9_728 papers). Shares no paper with JackHsieh/statML-arxiv-40M-20M's test split (4_864 papers). Shares no paper with /work/nvme/bhwe/jhsieh1/runs/statml-nested/160M's train split (38_912 papers). Rows are stored in draw order and each window is seeded on its paper, so same-seed builds at different sizes nest: a smaller one's rows appear here byte for byte. Same schema and recipe as
JackHsieh/statML-arxiv-40M-20M.
| Split | Documents | Tokens |
|---|---|---|
train |
4_096 | 16_777_216 |
- Downloads last month
- 8