NOSA: Native and Offloadable Sparse Attention
Paper • 2510.13602 • Published • 7
Datasets:
bench stringclasses 8
values | idx int64 0 996 | sample int64 0 7 | meta stringlengths 9 86 | prompt_tokens int64 16 1.84k | completion_tokens int64 693 32.7k | total_tokens int64 898 32.7k | think_tokens int64 400 32.7k | answer_tokens int64 0 15.9k | finish_reason stringclasses 2
values | truncated bool 2
classes | token_ids listlengths 693 32.7k | text stringlengths 1.04k 143k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
aime24 | 0 | 0 | {"ID": "2024-II-4"} | 154 | 3,984 | 4,138 | 2,802 | 1,180 | stop | false | [
151667,
198,
32313,
11,
773,
358,
1184,
311,
11625,
419,
1849,
315,
89936,
21220,
37906,
448,
7332,
856,
11,
379,
11,
323,
1147,
13,
576,
3491,
6696,
752,
2326,
37906,
1447,
16,
13,
1487,
31807,
2075,
11884,
18348,
593,
284,
220,
16,
... | <think>
Okay, so I need to solve this system of logarithmic equations with variables x, y, and z. The problem gives me three equations:
1. log₂(x/(yz)) = 1/2
2. log₂(y/(xz)) = 1/3
3. log₂(z/(xy)) = 1/4
And I need to find the absolute value of log₂(x⁴y³z²), which is |log₂(x⁴y³z²)|, and express it as m/n where m and n ... |
aime24 | 0 | 1 | {"ID": "2024-II-4"} | 154 | 13,730 | 13,884 | 12,535 | 1,193 | stop | false | [
151667,
198,
32313,
11,
773,
358,
1184,
311,
11625,
419,
1849,
315,
89936,
21220,
37906,
448,
7332,
856,
11,
379,
11,
323,
1147,
13,
576,
3491,
6696,
752,
2326,
37906,
1447,
16,
13,
1487,
31807,
2075,
11884,
18348,
593,
284,
220,
16,
... | <think>
Okay, so I need to solve this system of logarithmic equations with variables x, y, and z. The problem gives me three equations:
1. log₂(x/(yz)) = 1/2
2. log₂(y/(xz)) = 1/3
3. log₂(z/(xy)) = 1/4
And I need to find the value of |log₂(x⁴y³z²)|, which is m/n where m and n are coprime, then find m + n.
First, I r... |
aime24 | 0 | 2 | {"ID": "2024-II-4"} | 154 | 5,025 | 5,179 | 3,670 | 1,353 | stop | false | [
151667,
198,
32313,
11,
773,
358,
1184,
311,
11625,
419,
1849,
315,
89936,
21220,
37906,
448,
7332,
856,
11,
379,
11,
323,
1147,
13,
576,
3491,
6696,
752,
2326,
37906,
1447,
16,
13,
1487,
31807,
2075,
11884,
18348,
593,
284,
220,
16,
... | <think>
Okay, so I need to solve this system of logarithmic equations with variables x, y, and z. The problem gives me three equations:
1. log₂(x/(yz)) = 1/2
2. log₂(y/(xz)) = 1/3
3. log₂(z/(xy)) = 1/4
And I need to find the absolute value of log₂(x⁴y³z²), which is |log₂(x⁴y³z²)|, and express it as m/n where m and n ... |
aime24 | 0 | 3 | {"ID": "2024-II-4"} | 154 | 11,929 | 12,083 | 11,042 | 885 | stop | false | [151667,198,32313,11,773,358,1184,311,11625,419,1849,315,89936,21220,37906,448,7332,856,11,379,11,11(...TRUNCATED) | "<think>\nOkay, so I need to solve this system of logarithmic equations with variables x, y, z. The (...TRUNCATED) |
aime24 | 0 | 4 | {"ID": "2024-II-4"} | 154 | 10,570 | 10,724 | 9,436 | 1,132 | stop | false | [151667,198,32313,11,1077,594,1490,13,358,1184,311,11625,419,1849,315,89936,21220,37906,323,1477,279(...TRUNCATED) | "<think>\nOkay, let's see. I need to solve this system of logarithmic equations and find the value o(...TRUNCATED) |
aime24 | 0 | 5 | {"ID": "2024-II-4"} | 154 | 5,707 | 5,861 | 4,385 | 1,320 | stop | false | [151667,198,32313,11,773,358,1184,311,11625,419,1849,315,89936,21220,37906,448,7332,856,11,379,11,32(...TRUNCATED) | "<think>\nOkay, so I need to solve this system of logarithmic equations with variables x, y, and z. (...TRUNCATED) |
aime24 | 0 | 6 | {"ID": "2024-II-4"} | 154 | 7,084 | 7,238 | 5,969 | 1,113 | stop | false | [151667,198,32313,11,773,358,1184,311,11625,419,1849,315,89936,21220,37906,448,7332,856,11,379,11,32(...TRUNCATED) | "<think>\nOkay, so I need to solve this system of logarithmic equations with variables x, y, and z. (...TRUNCATED) |
aime24 | 0 | 7 | {"ID": "2024-II-4"} | 154 | 5,908 | 6,062 | 4,385 | 1,521 | stop | false | [151667,198,32313,11,773,358,1184,311,11625,419,1849,315,89936,21220,37906,323,1477,279,897,315,760,(...TRUNCATED) | "<think>\nOkay, so I need to solve this system of logarithmic equations and find the value of |log(...TRUNCATED) |
aime24 | 1 | 0 | {"ID": "2024-II-12"} | 186 | 23,222 | 23,408 | 21,720 | 1,500 | stop | false | [151667,198,32313,11,773,358,1184,311,11625,419,17047,3491,13,6771,752,1430,311,3535,432,3019,553,30(...TRUNCATED) | "<think>\nOkay, so I need to solve this geometry problem. Let me try to understand it step by step. (...TRUNCATED) |
aime24 | 1 | 1 | {"ID": "2024-II-12"} | 186 | 22,425 | 22,611 | 20,529 | 1,894 | stop | false | [151667,198,32313,11,773,358,1184,311,11625,419,17047,3491,13,6771,752,1430,311,3535,432,3019,553,30(...TRUNCATED) | "<think>\nOkay, so I need to solve this geometry problem. Let me try to understand it step by step. (...TRUNCATED) |
End of preview. Expand in Data Studio
Qwen3-1.7B reasoning traces
On-policy chain-of-thought rollouts from Qwen/Qwen3-1.7B
in thinking mode, collected to study long-context KV-cache residency constraints (trainable
sparse attention, in the InfLLM-V2 /
NOSA line).
Used as adaptation data: training a sparse+local attention variant on the model's own output distribution avoids the alignment tax that continued pretraining on external corpora imposes on a post-trained model.
Caveats
- Not filtered for correctness. Answers are unverified and many are wrong. The purpose is on-distribution rollout generation, not solution quality. Do not treat these as ground truth.
- Model-generated text: expect reasoning errors, dead ends, and self-contradiction. That is part of the distribution being modelled.
- Sampling:
temperature=0.6, top_p=0.95, top_k=20(Qwen3's official thinking-mode settings),max_model_len=32768,max_tokensset per prompt to32768 - prompt_tokens - 64so traces run to natural length rather than a fixed cap.
Source benchmarks
Only the problem statements are used.
bench |
source | items x samples |
|---|---|---|
aime24 |
Maxwell-Jia/AIME_2024 |
30 x 8 |
aime25 |
yentinglin/aime_2025 |
30 x 8 |
math500 |
HuggingFaceH4/MATH-500 |
500 x 1 |
zebralogic |
allenai/ZebraLogicBench (grid_mode) |
200 x 1 |
autologi |
qzhu/AutoLogi (English subset) |
200 x 1 |
Schema
bench, idx, sample, text (full completion incl. <think>...</think>), token_ids
(Qwen3 tokenizer), prompt_tokens, completion_tokens, total_tokens, think_tokens,
answer_tokens, finish_reason, truncated, meta (JSON string of benchmark-specific fields).
from datasets import load_dataset
ds = load_dataset("kmoss/qwen3-1.7b-traces", split="train")
- Downloads last month
- 86