Dataset Viewer
Auto-converted to Parquet Duplicate
split
stringclasses
4 values
question_id
stringlengths
26
43
sample_index
int16
0
15
text
stringlengths
103
7.74k
⌀
token_ids
listlengths
31
2.05k
⌀
sampled_logprobs
listlengths
31
2.05k
⌀
self_certainty
listlengths
31
2.05k
⌀
finish_reason
stringclasses
2 values
truncated
bool
2 classes
generated_token_count
int32
31
2.05k
⌀
reference_chat_tokens
int32
57
2.99k
⌀
request_seed
int64
12.8k
2.15B
⌀
generation_identity
stringclasses
4 values
native_generation_seconds_share
float64
0.06
4.5
⌀
lowest_confidence
float64
0.71
28.9
⌀
bottom10_confidence
float64
2.98
32
⌀
extraction_inline_input_tokens
int32
84
3.02k
⌀
extraction_inline_output_tokens
int32
1
128
⌀
extraction_inline_seconds_share
float64
0
0.31
⌀
extraction_final_input_tokens
int32
84
3.02k
⌀
extraction_final_output_tokens
int32
1
128
⌀
extraction_final_seconds_share
float64
0
0.29
⌀
pred
stringlengths
0
574
⌀
prediction
stringlengths
0
475
⌀
extraction_failed
bool
2 classes
embedding_valid
bool
1 class
reasoning_mean_logprob
float64
-0.66
-0
⌀
answer_mean_logprob
float64
-6.94
0
⌀
reasoning_mean_entropy
float64
0.01
0.76
⌀
reasoning_token_count
int32
30
2.05k
⌀
answer_token_count
int32
0
125
⌀
feature_input_tokens
int32
56
2.99k
⌀
feature_seconds_share
float64
0.01
0.21
⌀
correct
bool
2 classes
raw_correct
bool
2 classes
calibration
study_gsm8k_calibration_4a043e8d63be6481
0
Let the total number of seats in the auditorium be $S = 500$. The number of seats taken is $\frac{2}{5}$ of the total number of seats. So, the number of seats taken is $$ \frac{2}{5} \times 500 = 2 \times 100 = 200 $$ The number of seats broken is $\frac{1}{10}$ of the total number of seats. So, the number of broken se...
[ 6481, 506, 2558, 1548, 529, 17823, 528, 506, 104569, 577, 609, 236773, 578, 236743, 236810, 236771, 236771, 2104, 107, 818, 1548, 529, 17823, 3523, 563, 874, 1420, 236782, 236778, 1170, 236810, 823, 529, 506, 2558, 1548, 529, 17823, 236761,...
[ -0.00012957210128661245, -0.018170330673456192, -0.00002312633478140924, -0.0011717366287484765, 0, -0.000004768360213347478, -1.1920928244535389e-7, -0.000006198863957251888, 0, -4.768370445162873e-7, -0.12708280980587006, -1.1701924800872803, -0.023690031841397285, 0, 0, 0, 0, -0...
[ 12.085647583007812, 13.622437477111816, 15.973040580749512, 18.44293975830078, 27.17926788330078, 20.6292724609375, 21.298017501831055, 21.263648986816406, 21.754268646240234, 17.079267501831055, 13.325100898742676, 13.38071060180664, 15.12170696258545, 39.091766357421875, 34.92301940917...
stop
false
626
678
578,953,685
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.159544
14.648719
15.580715
701
5
0.002346
701
5
0.002367
250
250
false
true
-0.093768
0
0.092092
619
3
677
0.02142
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
1
Let the total number of seats in the auditorium be $T = 500$. The number of seats taken is $\frac{2}{5}$ of the total seats, so the number of seats taken is $\frac{2}{5} \times 500 = 2 \times 100 = 200$. The number of seats broken is $\frac{1}{10}$ of the total seats, so the number of seats broken is $\frac{1}{10} \tim...
[ 6481, 506, 2558, 1548, 529, 17823, 528, 506, 104569, 577, 609, 236774, 578, 236743, 236810, 236771, 236771, 2104, 107, 818, 1548, 529, 17823, 3523, 563, 874, 1420, 236782, 236778, 1170, 236810, 823, 529, 506, 2558, 17823, 236764, 834, 506, ...
[ -0.00012957210128661245, -0.018170330673456192, -0.00002312633478140924, -0.0011717366287484765, 0, -0.000004768360213347478, -1.1920928244535389e-7, -0.000006198863957251888, 0, -4.768370445162873e-7, -0.12708280980587006, -0.4201924502849579, -0.03859463334083557, 0, 0, 0, 0, -0....
[ 12.085647583007812, 13.622437477111816, 15.973040580749512, 18.44293975830078, 27.17926788330078, 20.6292724609375, 21.298017501831055, 21.263648986816406, 21.754268646240234, 17.079267501831055, 13.325100898742676, 13.38071060180664, 15.911612510681152, 39.54801940917969, 34.84176635742...
stop
false
535
587
1,891,665,999
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
15.642352
16.149714
610
5
0.002346
610
5
0.002367
250
250
false
true
-0.081702
0
0.098078
528
3
586
0.017548
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
2
Let the total number of seats in the auditorium be $S = 500$. The number of seats taken is $\frac{2}{5}$ of the total seats, so the number of seats taken is $\frac{2}{5} \times 500 = 2 \times 100 = 200$. The number of seats broken is $\frac{1}{10}$ of the total seats, so the number of broken seats is $\frac{1}{10} \tim...
[ 6481, 506, 2558, 1548, 529, 17823, 528, 506, 104569, 577, 609, 236773, 578, 236743, 236810, 236771, 236771, 2104, 107, 818, 1548, 529, 17823, 3523, 563, 874, 1420, 236782, 236778, 1170, 236810, 823, 529, 506, 2558, 17823, 236764, 834, 506, ...
[ -0.00012957210128661245, -0.018170330673456192, -0.00002312633478140924, -0.0011717366287484765, 0, -0.000004768360213347478, -1.1920928244535389e-7, -0.000006198863957251888, 0, -4.768370445162873e-7, -0.12708280980587006, -1.1701924800872803, -0.023690031841397285, 0, 0, 0, 0, -0...
[ 12.085647583007812, 13.622437477111816, 15.973040580749512, 18.44293975830078, 27.17926788330078, 20.6292724609375, 21.298017501831055, 21.263648986816406, 21.754268646240234, 17.079267501831055, 13.325100898742676, 13.38071060180664, 15.12170696258545, 39.091766357421875, 34.92301940917...
stop
false
642
694
1,575,098,859
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
15.840405
16.53667
717
5
0.002346
717
5
0.002367
250
250
false
true
-0.087861
-0.000002
0.100824
635
3
693
0.02142
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
3
Let the total number of seats in the auditorium be $T = 500$. The number of seats taken is $\frac{2}{5}$ of the total number of seats, so the number of seats taken is $\frac{2}{5} \times 500 = 2 \times 100 = 200$. The number of seats broken is $\frac{1}{10}$ of the total number of seats, so the number of broken seats i...
[ 6481, 506, 2558, 1548, 529, 17823, 528, 506, 104569, 577, 609, 236774, 578, 236743, 236810, 236771, 236771, 2104, 107, 818, 1548, 529, 17823, 3523, 563, 874, 1420, 236782, 236778, 1170, 236810, 823, 529, 506, 2558, 1548, 529, 17823, 236764,...
[ -0.00012957210128661245, -0.018170330673456192, -0.00002312633478140924, -0.0011717366287484765, 0, -0.000004768360213347478, -1.1920928244535389e-7, -0.000006198863957251888, 0, -4.768370445162873e-7, -0.12708280980587006, -0.4201924502849579, -0.03859463334083557, 0, 0, 0, 0, -0....
[ 12.085647583007812, 13.622437477111816, 15.973040580749512, 18.44293975830078, 27.17926788330078, 20.6292724609375, 21.298017501831055, 21.263648986816406, 21.754268646240234, 17.079267501831055, 13.325100898742676, 13.38071060180664, 15.911612510681152, 39.54801940917969, 34.84176635742...
stop
false
444
496
1,491,205,372
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
15.06265
15.574369
519
5
0.002346
519
5
0.002367
250
250
false
true
-0.088133
0
0.084907
437
3
495
0.014339
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
4
Let the total number of seats in the auditorium be $T = 500$. The number of seats taken is $\frac{2}{5}$ of the total seats, so the number of seats taken is $\frac{2}{5} \times 500 = 2 \times 100 = 200$. The number of seats broken is $\frac{1}{10}$ of the total seats, so the number of broken seats is $\frac{1}{10} \tim...
[ 6481, 506, 2558, 1548, 529, 17823, 528, 506, 104569, 577, 609, 236774, 578, 236743, 236810, 236771, 236771, 2104, 107, 818, 1548, 529, 17823, 3523, 563, 874, 1420, 236782, 236778, 1170, 236810, 823, 529, 506, 2558, 17823, 236764, 834, 506, ...
[ -0.00012957210128661245, -0.018170330673456192, -0.00002312633478140924, -0.0011717366287484765, 0, -0.000004768360213347478, -1.1920928244535389e-7, -0.000006198863957251888, 0, -4.768370445162873e-7, -0.12708280980587006, -0.4201924502849579, -0.03859463334083557, 0, 0, 0, 0, -0....
[ 12.085647583007812, 13.622437477111816, 15.973040580749512, 18.44293975830078, 27.17926788330078, 20.6292724609375, 21.298017501831055, 21.263648986816406, 21.754268646240234, 17.079267501831055, 13.325100898742676, 13.38071060180664, 15.911612510681152, 39.54801940917969, 34.84176635742...
stop
false
436
488
1,340,297,573
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
17.478087
17.73678
511
5
0.002346
511
5
0.002367
250
250
false
true
-0.082295
0
0.086443
429
3
487
0.01496
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
5
Let the total number of seats in the auditorium be 500. The number of seats taken is $\frac{2}{5}$ of the total number of seats, so the number of seats taken is $\frac{2}{5} \times 500$. $\frac{2}{5} \times 500 = 2 \times 100 = 200$ seats are taken. The number of seats broken is $\frac{1}{10}$ of the total number of se...
[ 6481, 506, 2558, 1548, 529, 17823, 528, 506, 104569, 577, 236743, 236810, 236771, 236771, 236761, 107, 818, 1548, 529, 17823, 3523, 563, 874, 1420, 236782, 236778, 1170, 236810, 823, 529, 506, 2558, 1548, 529, 17823, 236764, 834, 506, 1548,...
[ -0.00012957210128661245, -0.018170330673456192, -0.00002312633478140924, -0.0011717366287484765, 0, -0.000004768360213347478, -1.1920928244535389e-7, -0.000006198863957251888, 0, -4.768370445162873e-7, -2.1270828247070312, 0, 0, 0, -0.000012993727978027891, -0.01416497677564621, -0.0...
[ 12.085647583007812, 13.622437477111816, 15.973040580749512, 18.44293975830078, 27.17926788330078, 20.6292724609375, 21.298017501831055, 21.263648986816406, 21.754268646240234, 17.079267501831055, 13.325100898742676, 23.541767120361328, 24.729267120361328, 22.80426788330078, 20.2980308532...
stop
false
875
927
819,359,917
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
7.557824
8.295421
950
5
0.002346
950
5
0.002367
250
250
false
true
-0.12401
-0.000017
0.147818
868
3
926
0.026366
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
6
"Let the total number of seats in the auditorium be $T = 500$.\nThe number of seats taken is $\\frac(...TRUNCATED)
[6481,506,2558,1548,529,17823,528,506,104569,577,609,236774,578,236743,236810,236771,236771,2104,107(...TRUNCATED)
[-0.00012957210128661245,-0.018170330673456192,-0.00002312633478140924,-0.0011717366287484765,0.0,-4(...TRUNCATED)
[12.085647583007812,13.622437477111816,15.973040580749512,18.44293975830078,27.17926788330078,20.629(...TRUNCATED)
length
true
1,024
1,077
481,118,195
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
10.058181
10.821692
1,100
5
0.002346
1,100
5
0.002367
250
250
true
true
-0.12939
0
0.148833
1,024
0
1,075
0.030155
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
7
"Let the total number of seats in the auditorium be $S = 500$.\nThe number of seats that are current(...TRUNCATED)
[6481,506,2558,1548,529,17823,528,506,104569,577,609,236773,578,236743,236810,236771,236771,2104,107(...TRUNCATED)
[-0.00012957210128661245,-0.018170330673456192,-0.00002312633478140924,-0.0011717366287484765,0.0,-4(...TRUNCATED)
[12.085647583007812,13.622437477111816,15.973040580749512,18.44293975830078,27.17926788330078,20.629(...TRUNCATED)
stop
false
672
724
1,289,921,164
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
12.567887
13.137175
747
5
0.002346
747
5
0.002367
250
250
false
true
-0.098157
-0
0.099381
665
3
723
0.02142
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
8
"Let the total number of seats in the auditorium be $T = 500$.\nThe number of seats that are taken i(...TRUNCATED)
[6481,506,2558,1548,529,17823,528,506,104569,577,609,236774,578,236743,236810,236771,236771,2104,107(...TRUNCATED)
[-0.00012957210128661245,-0.018170330673456192,-0.00002312633478140924,-0.0011717366287484765,0.0,-4(...TRUNCATED)
[12.085647583007812,13.622437477111816,15.973040580749512,18.44293975830078,27.17926788330078,20.629(...TRUNCATED)
stop
false
431
483
1,154,948,637
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
16.847545
17.542552
506
5
0.002346
506
5
0.002367
250
250
false
true
-0.071741
0
0.081601
424
3
482
0.01496
true
true
calibration
study_gsm8k_calibration_4a043e8d63be6481
9
"Let the total number of seats in the auditorium be $T = 500$.\nThe number of seats taken is $\\frac(...TRUNCATED)
[6481,506,2558,1548,529,17823,528,506,104569,577,609,236774,578,236743,236810,236771,236771,2104,107(...TRUNCATED)
[-0.00012957210128661245,-0.018170330673456192,-0.00002312633478140924,-0.0011717366287484765,0.0,-4(...TRUNCATED)
[12.085647583007812,13.622437477111816,15.973040580749512,18.44293975830078,27.17926788330078,20.629(...TRUNCATED)
stop
false
548
600
1,247,794,940
67b4f7efa6a927768d976765a81a1e612321f23eadd4f1e5bc4b904b29ecdb3c
0.127802
15.609575
16.250642
623
5
0.002346
623
5
0.002367
250
250
false
true
-0.093553
0
0.089762
541
3
599
0.019487
true
true
End of preview. Expand in Data Studio

PGM adaptive self-consistency: released artifacts

Data and fitted models for github.com/Ever2after/PGM, which compares a partition-mixture aggregation head (PGM) for adaptive self-consistency with matched ablations and the SC / ASC / ESC / ReASC baselines on MATH500, GSM8K and Omni-MATH with Gemma-3-4B-IT, Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct.

Folder Files Content
cohorts/ 3 Questions and gold answers per dataset, splits calibration / train / dev / test
banks/ 36 Per (model, dataset): 16 sampled responses per question with teacher-forcing features, extraction costs, ReASC self-certainty and xVerify-3B-Ia labels (.parquet), plus row-aligned reasoning embeddings (.embeddings.npy, float32) and per-token top-20 log-probabilities (.top_logprobs.parquet, 18 GB in total, optional)
policies/ 24 Per (model, dataset): fitted policy (.json: partition thresholds, scalers, temperatures, dev-selected thresholds, baseline calibration) and the 36 head checkpoints (.checkpoints.pt)
decisions/ 16 Per evaluation cell: the decision of every method on every test question
manifest.json 1 Size and SHA256 of every file

396,928 responses in total (12 banks of 26,048–39,152 responses).

Use

git clone https://github.com/Ever2after/PGM && cd PGM
pip install -e .
pgm download                 # everything except log-probabilities (~7 GB); or e.g. --only cohorts policies --models qwen7b
pgm download --only logprobs  # per-token top-20 log-probabilities (18 GB)
pgm evaluate all             # replays every reported number on CPU

Without the code:

import numpy as np, pyarrow.parquet as pq
bank = pq.read_table("banks/qwen7b_math.parquet")
emb = np.load("banks/qwen7b_math.embeddings.npy", mmap_mode="r")   # emb[i] belongs to bank row i

Column definitions and the policy format are documented in docs/DATA.md.

Provenance and licenses

Responses were sampled with vLLM 0.11.0 from the pinned model revisions in configs/models.json using each model's official sampling defaults; labels come from the pinned public xVerify-3B-Ia judge. The questions come from MATH (Hendrycks et al.; MATH500 subset of OpenAI PRM800K), GSM8K (OpenAI) and Omni-MATH (KbsdJames/Omni-MATH), and remain under their original licenses. Generated responses are subject to the terms of the respective model licenses (Gemma Terms of Use, Qwen license, Llama 3.1 Community License). This release adds no rights to the upstream datasets or models.

The complete original run archive (including earlier V2/V5 experiments) is in Ever2after/PGM-experiment-artifacts.

Downloads last month
87