Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
44
55
pair_id
stringlengths
29
36
paper_id
stringlengths
25
32
forum_id
stringclasses
41 values
split
stringclasses
1 value
variant
class label
2 classes
title
stringlengths
13
153
authors
listlengths
1
13
conference
stringclasses
3 values
year
int32
2.01k
2.02k
source_page
int32
1
10
pdf_url
stringlengths
37
101
abstract_url
stringlengths
34
105
pdf_sha256
stringlengths
64
64
development_exposed
bool
2 classes
eligible_for_pilot_training
bool
2 classes
quality_status
stringclasses
2 values
quality_flags
listlengths
0
7
generation_id
stringlengths
35
35
⌀
generator_model
stringclasses
1 value
generator_canonical_model
stringclasses
1 value
generation_service_tier
stringclasses
2 values
fidelity_verdict
stringclasses
5 values
paired_generation_fidelity_verdict
stringclasses
4 values
paired_quality_audit_json
stringlengths
169
1.07k
paired_fidelity_review_json
stringlengths
409
2.43k
sample_weight
float64
1
1
text
stringlengths
533
5.84k
text_sha256
stringlengths
64
64
source_text_sha256
stringlengths
64
64
label
class label
2 classes
has_ai
bool
2 classes
target_start
int32
61
3.66k
target_end
int32
327
4.49k
target_label
class label
2 classes
offset_unit
stringclasses
1 value
offset_convention
stringclasses
1 value
regions
listlengths
1
3
input_ids
listlengths
115
1.24k
attention_mask
listlengths
115
1.24k
token_labels
listlengths
115
1.24k
token_loss_mask
listlengths
115
1.24k
token_start
listlengths
115
1.24k
token_end
listlengths
115
1.24k
token_text
listlengths
115
1.24k
tokenizer_sha256
stringclasses
1 value
words
listlengths
84
904
word_labels
listlengths
84
904
word_start
listlengths
84
904
word_end
listlengths
84
904
sentences
listlengths
3
45
723e8f97fde15f7a8d5ff8d558ea3f16/g01/human_original
723e8f97fde15f7a8d5ff8d558ea3f16/g01
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
0human_original
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
2
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
null
null
null
null
not_applicable
fully_faithful
{"citations_preserved":true,"clarity":"better","content_covered":true,"context_fit":"good","issues":[],"meaning_preserved":true,"outline_faithful":true,"source_issues":[]}
{"differences":[],"dimensions":{"attribution_citations":"preserved","explicit_content":"preserved","implications":"preserved","qualifications":"preserved","quantitative_details":"not_applicable","tone_rhetorical_function":"preserved"},"explanation":"The candidate preserves both cited works’ methods and claims, includin...
1
Existing work either models the whole drawing (as an image) or as a complete sequence of points. In contrast, we model stroke-based drawings as order invariant 2D compositional structures and in consequence our model scales to more complex settings. Albeit in diverse and different domains, the following works are also ...
fa5e83bbc1fe5fc96e7fcf27642ef3908c22255ebe5624127165cf90af7dff8a
fa5e83bbc1fe5fc96e7fcf27642ef3908c22255ebe5624127165cf90af7dff8a
0human
false
415
993
0human
unicode_code_points
half_open
[ { "start": 0, "end": 1273, "label": 0, "source_label": "human_preserved", "source_start": 0, "source_end": 1273 } ]
[ 1672, 9020, 789, 2057, 3210, 253, 2644, 10263, 313, 284, 271, 2460, 10, 390, 347, 247, 3426, 3425, 273, 2792, 15, 496, 4499, 13, 359, 1566, 9980, 14, 3169, 21780, 347, 1340, 13727, 374, 37, 5889, 267, 5289, 285, 275, 9936, 776, 1566...
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true...
[ 0, 2, 8, 13, 20, 27, 31, 37, 45, 47, 49, 52, 58, 59, 62, 65, 67, 76, 85, 88, 95, 96, 99, 108, 109, 112, 118, 125, 126, 131, 140, 143, 149, 159, 161, 162, 174, 176, 187, 191, 194, 206, 210, 216, 223, 226, 231, 239, 248, ...
[ 2, 8, 13, 20, 27, 31, 37, 45, 47, 49, 52, 58, 59, 62, 65, 67, 76, 85, 88, 95, 96, 99, 108, 109, 112, 118, 125, 126, 131, 140, 143, 149, 159, 161, 162, 174, 176, 187, 191, 194, 206, 210, 216, 223, 226, 231, 239, 248, 249,...
[ "Ex", "isting", " work", " either", " models", " the", " whole", " drawing", " (", "as", " an", " image", ")", " or", " as", " a", " complete", " sequence", " of", " points", ".", " In", " contrast", ",", " we", " model", " stroke", "-", "based", " drawings"...
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
[ "Existing", "work", "either", "models", "the", "whole", "drawing", "(as", "an", "image)", "or", "as", "a", "complete", "sequence", "of", "points.", "In", "contrast,", "we", "model", "stroke-based", "drawings", "as", "order", "invariant", "2D", "compositional", ...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ 0, 9, 14, 21, 28, 32, 38, 46, 50, 53, 60, 63, 66, 68, 77, 86, 89, 97, 100, 110, 113, 119, 132, 141, 144, 150, 160, 163, 177, 188, 192, 195, 207, 211, 217, 224, 227, 232, 240, 250, 257, 260, 268, 272, 282, 291, 295, 305, ...
[ 8, 13, 20, 27, 31, 37, 45, 49, 52, 59, 62, 65, 67, 76, 85, 88, 96, 99, 109, 112, 118, 131, 140, 143, 149, 159, 162, 176, 187, 191, 194, 206, 210, 216, 223, 226, 231, 239, 249, 256, 259, 267, 271, 281, 290, 294, 304, 310, ...
[ { "start": 0, "end": 96, "text": "Existing work either models the whole drawing (as an image) or as a complete sequence of points.", "label": 0, "binary_label": 0, "loss_mask": true, "ai_character_fraction": 0, "human_characters": 96, "ai_characters": 0 }, { "start": 97, ...
723e8f97fde15f7a8d5ff8d558ea3f16/g01/paragraph_generate
723e8f97fde15f7a8d5ff8d558ea3f16/g01
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
1paragraph_generated
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
2
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
gen-1790736978-ryRl9ypsibTRem0pguh7
openai/gpt-6-luna
openai/gpt-6-luna-20260922
default
fully_faithful
fully_faithful
{"citations_preserved":true,"clarity":"better","content_covered":true,"context_fit":"good","issues":[],"meaning_preserved":true,"outline_faithful":true,"source_issues":[]}
{"differences":[],"dimensions":{"attribution_citations":"preserved","explicit_content":"preserved","implications":"preserved","qualifications":"preserved","quantitative_details":"not_applicable","tone_rhetorical_function":"preserved"},"explanation":"The candidate preserves both cited works’ methods and claims, includin...
1
Existing work either models the whole drawing (as an image) or as a complete sequence of points. In contrast, we model stroke-based drawings as order invariant 2D compositional structures and in consequence our model scales to more complex settings. Albeit in diverse and different domains, the following works are also ...
c954977a75e156f003b325b0cf2617a7a3464931b414fe666c33b7913f67c5f4
fa5e83bbc1fe5fc96e7fcf27642ef3908c22255ebe5624127165cf90af7dff8a
2mixed
true
415
969
1ai
unicode_code_points
half_open
[ { "start": 0, "end": 415, "label": 0, "source_label": "human_preserved", "source_start": 0, "source_end": 415 }, { "start": 415, "end": 969, "label": 1, "source_label": "ai_rewritten", "source_start": 415, "source_end": 993 }, { "start": 969, "end": 12...
[ 1672, 9020, 789, 2057, 3210, 253, 2644, 10263, 313, 284, 271, 2460, 10, 390, 347, 247, 3426, 3425, 273, 2792, 15, 496, 4499, 13, 359, 1566, 9980, 14, 3169, 21780, 347, 1340, 13727, 374, 37, 5889, 267, 5289, 285, 275, 9936, 776, 1566...
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true...
[ 0, 2, 8, 13, 20, 27, 31, 37, 45, 47, 49, 52, 58, 59, 62, 65, 67, 76, 85, 88, 95, 96, 99, 108, 109, 112, 118, 125, 126, 131, 140, 143, 149, 159, 161, 162, 174, 176, 187, 191, 194, 206, 210, 216, 223, 226, 231, 239, 248, ...
[ 2, 8, 13, 20, 27, 31, 37, 45, 47, 49, 52, 58, 59, 62, 65, 67, 76, 85, 88, 95, 96, 99, 108, 109, 112, 118, 125, 126, 131, 140, 143, 149, 159, 161, 162, 174, 176, 187, 191, 194, 206, 210, 216, 223, 226, 231, 239, 248, 249,...
[ "Ex", "isting", " work", " either", " models", " the", " whole", " drawing", " (", "as", " an", " image", ")", " or", " as", " a", " complete", " sequence", " of", " points", ".", " In", " contrast", ",", " we", " model", " stroke", "-", "based", " drawings"...
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
[ "Existing", "work", "either", "models", "the", "whole", "drawing", "(as", "an", "image)", "or", "as", "a", "complete", "sequence", "of", "points.", "In", "contrast,", "we", "model", "stroke-based", "drawings", "as", "order", "invariant", "2D", "compositional", ...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ 0, 9, 14, 21, 28, 32, 38, 46, 50, 53, 60, 63, 66, 68, 77, 86, 89, 97, 100, 110, 113, 119, 132, 141, 144, 150, 160, 163, 177, 188, 192, 195, 207, 211, 217, 224, 227, 232, 240, 250, 257, 260, 268, 272, 282, 291, 295, 305, ...
[ 8, 13, 20, 27, 31, 37, 45, 49, 52, 59, 62, 65, 67, 76, 85, 88, 96, 99, 109, 112, 118, 131, 140, 143, 149, 159, 162, 176, 187, 191, 194, 206, 210, 216, 223, 226, 231, 239, 249, 256, 259, 267, 271, 281, 290, 294, 304, 310, ...
[ { "start": 0, "end": 96, "text": "Existing work either models the whole drawing (as an image) or as a complete sequence of points.", "label": 0, "binary_label": 0, "loss_mask": true, "ai_character_fraction": 0, "human_characters": 96, "ai_characters": 0 }, { "start": 97, ...
723e8f97fde15f7a8d5ff8d558ea3f16/g02/human_original
723e8f97fde15f7a8d5ff8d558ea3f16/g02
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
0human_original
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
2
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
null
null
null
null
not_applicable
fully_faithful
{"citations_preserved":true,"clarity":"better","content_covered":true,"context_fit":"good","issues":[],"meaning_preserved":true,"outline_faithful":true,"source_issues":[]}
{"differences":[],"dimensions":{"attribution_citations":"not_applicable","explicit_content":"preserved","implications":"preserved","qualifications":"preserved","quantitative_details":"not_applicable","tone_rhetorical_function":"preserved"},"explanation":"The candidate preserves the contrast with prior approaches, the c...
1
Particularly relevant to our work are approaches that apply generative models to stroke data. Ha et. al. [6] and Ribeiro et. al. [7] build LSTM/VAE-based and Transformer-based models respectively to generate samples from the QuickDraw dataset [20]. These approaches model the entire drawing as a single sequence of point...
b29b6e9899cb73e0f0e7707e27c5834f073f2953e630e94520ad9766be4aa372
b29b6e9899cb73e0f0e7707e27c5834f073f2953e630e94520ad9766be4aa372
0human
false
818
1,231
0human
unicode_code_points
half_open
[ { "start": 0, "end": 1811, "label": 0, "source_label": "human_preserved", "source_start": 0, "source_end": 1811 } ]
[ 7834, 21883, 4623, 281, 776, 789, 403, 7274, 326, 4647, 1006, 800, 3210, 281, 9980, 941, 15, 10664, 1162, 15, 355, 15, 544, 23, 62, 285, 416, 24669, 9401, 1162, 15, 355, 15, 544, 24, 62, 1973, 418, 27127, 16, 55, 12648, 14, 3169, ...
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true...
[ 0, 4, 12, 21, 24, 28, 33, 37, 48, 53, 59, 65, 70, 77, 80, 87, 92, 93, 96, 99, 100, 103, 104, 106, 107, 108, 112, 114, 117, 120, 123, 124, 127, 128, 130, 131, 132, 138, 140, 143, 144, 145, 147, 148, 153, 157, 163, 169, 17...
[ 4, 12, 21, 24, 28, 33, 37, 48, 53, 59, 65, 70, 77, 80, 87, 92, 93, 96, 99, 100, 103, 104, 106, 107, 108, 112, 114, 117, 120, 123, 124, 127, 128, 130, 131, 132, 138, 140, 143, 144, 145, 147, 148, 153, 157, 163, 169, 170, ...
[ "Part", "icularly", " relevant", " to", " our", " work", " are", " approaches", " that", " apply", " gener", "ative", " models", " to", " stroke", " data", ".", " Ha", " et", ".", " al", ".", " [", "6", "]", " and", " R", "ibe", "iro", " et", ".", " al",...
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
[ "Particularly", "relevant", "to", "our", "work", "are", "approaches", "that", "apply", "generative", "models", "to", "stroke", "data.", "Ha", "et.", "al.", "[6]", "and", "Ribeiro", "et.", "al.", "[7]", "build", "LSTM/VAE-based", "and", "Transformer-based", "mode...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ 0, 13, 22, 25, 29, 34, 38, 49, 54, 60, 71, 78, 81, 88, 94, 97, 101, 105, 109, 113, 121, 125, 129, 133, 139, 154, 158, 176, 183, 196, 199, 208, 216, 221, 225, 235, 243, 249, 255, 266, 272, 276, 283, 291, 294, 296, 303, 312,...
[ 12, 21, 24, 28, 33, 37, 48, 53, 59, 70, 77, 80, 87, 93, 96, 100, 104, 108, 112, 120, 124, 128, 132, 138, 153, 157, 175, 182, 195, 198, 207, 215, 220, 224, 234, 242, 248, 254, 265, 271, 275, 282, 290, 293, 295, 302, 311, 31...
[ { "start": 0, "end": 93, "text": "Particularly relevant to our work are approaches that apply generative models to stroke data.", "label": 0, "binary_label": 0, "loss_mask": true, "ai_character_fraction": 0, "human_characters": 93, "ai_characters": 0 }, { "start": 94, ...
723e8f97fde15f7a8d5ff8d558ea3f16/g02/paragraph_generate
723e8f97fde15f7a8d5ff8d558ea3f16/g02
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
1paragraph_generated
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
2
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
gen-1790736978-GgXb3JdEm4WObUt71YgD
openai/gpt-6-luna
openai/gpt-6-luna-20260922
default
fully_faithful
fully_faithful
{"citations_preserved":true,"clarity":"better","content_covered":true,"context_fit":"good","issues":[],"meaning_preserved":true,"outline_faithful":true,"source_issues":[]}
{"differences":[],"dimensions":{"attribution_citations":"not_applicable","explicit_content":"preserved","implications":"preserved","qualifications":"preserved","quantitative_details":"not_applicable","tone_rhetorical_function":"preserved"},"explanation":"The candidate preserves the contrast with prior approaches, the c...
1
Particularly relevant to our work are approaches that apply generative models to stroke data. Ha et. al. [6] and Ribeiro et. al. [7] build LSTM/VAE-based and Transformer-based models respectively to generate samples from the QuickDraw dataset [20]. These approaches model the entire drawing as a single sequence of point...
71b01df5b8f86e6b992f38830868ce46f44d5a6a538b80084fc183ed8e698b46
b29b6e9899cb73e0f0e7707e27c5834f073f2953e630e94520ad9766be4aa372
2mixed
true
818
1,203
1ai
unicode_code_points
half_open
[ { "start": 0, "end": 818, "label": 0, "source_label": "human_preserved", "source_start": 0, "source_end": 818 }, { "start": 818, "end": 1203, "label": 1, "source_label": "ai_rewritten", "source_start": 818, "source_end": 1231 }, { "start": 1203, "end":...
[ 7834, 21883, 4623, 281, 776, 789, 403, 7274, 326, 4647, 1006, 800, 3210, 281, 9980, 941, 15, 10664, 1162, 15, 355, 15, 544, 23, 62, 285, 416, 24669, 9401, 1162, 15, 355, 15, 544, 24, 62, 1973, 418, 27127, 16, 55, 12648, 14, 3169, ...
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true...
[ 0, 4, 12, 21, 24, 28, 33, 37, 48, 53, 59, 65, 70, 77, 80, 87, 92, 93, 96, 99, 100, 103, 104, 106, 107, 108, 112, 114, 117, 120, 123, 124, 127, 128, 130, 131, 132, 138, 140, 143, 144, 145, 147, 148, 153, 157, 163, 169, 17...
[ 4, 12, 21, 24, 28, 33, 37, 48, 53, 59, 65, 70, 77, 80, 87, 92, 93, 96, 99, 100, 103, 104, 106, 107, 108, 112, 114, 117, 120, 123, 124, 127, 128, 130, 131, 132, 138, 140, 143, 144, 145, 147, 148, 153, 157, 163, 169, 170, ...
[ "Part", "icularly", " relevant", " to", " our", " work", " are", " approaches", " that", " apply", " gener", "ative", " models", " to", " stroke", " data", ".", " Ha", " et", ".", " al", ".", " [", "6", "]", " and", " R", "ibe", "iro", " et", ".", " al",...
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
[ "Particularly", "relevant", "to", "our", "work", "are", "approaches", "that", "apply", "generative", "models", "to", "stroke", "data.", "Ha", "et.", "al.", "[6]", "and", "Ribeiro", "et.", "al.", "[7]", "build", "LSTM/VAE-based", "and", "Transformer-based", "mode...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ 0, 13, 22, 25, 29, 34, 38, 49, 54, 60, 71, 78, 81, 88, 94, 97, 101, 105, 109, 113, 121, 125, 129, 133, 139, 154, 158, 176, 183, 196, 199, 208, 216, 221, 225, 235, 243, 249, 255, 266, 272, 276, 283, 291, 294, 296, 303, 312,...
[ 12, 21, 24, 28, 33, 37, 48, 53, 59, 70, 77, 80, 87, 93, 96, 100, 104, 108, 112, 120, 124, 128, 132, 138, 153, 157, 175, 182, 195, 198, 207, 215, 220, 224, 234, 242, 248, 254, 265, 271, 275, 282, 290, 293, 295, 302, 311, 31...
[ { "start": 0, "end": 93, "text": "Particularly relevant to our work are approaches that apply generative models to stroke data.", "label": 0, "binary_label": 0, "loss_mask": true, "ai_character_fraction": 0, "human_characters": 93, "ai_characters": 0 }, { "start": 94, ...
723e8f97fde15f7a8d5ff8d558ea3f16/g03/human_original
723e8f97fde15f7a8d5ff8d558ea3f16/g03
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
0human_original
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
2
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
null
null
null
null
not_applicable
fully_faithful
{"citations_preserved":true,"clarity":"better","content_covered":true,"context_fit":"good","issues":[],"meaning_preserved":true,"outline_faithful":true,"source_issues":[]}
{"differences":[],"dimensions":{"attribution_citations":"preserved","explicit_content":"preserved","implications":"preserved","qualifications":"preserved","quantitative_details":"preserved","tone_rhetorical_function":"preserved"},"explanation":"The candidate preserves the claims, examples, scope, number of diagrams, an...
1
The interpretation of stroke data has been pursued before deep learning, often on small datasets of a few hundred samples targeting a particular application: Costagliola et al. [14] presented a parsing-based approach using a grammar of shapes and symbols where shapes and symbols are independently recognized and the res...
8c38dcf2b223346655eb32b595cb9bd804dde2f234aed9cd45385f49dd2285ab
8c38dcf2b223346655eb32b595cb9bd804dde2f234aed9cd45385f49dd2285ab
0human
false
548
1,115
0human
unicode_code_points
half_open
[ { "start": 0, "end": 1933, "label": 0, "source_label": "human_preserved", "source_start": 0, "source_end": 1933 } ]
[ 510, 7914, 273, 9980, 941, 556, 644, 23321, 1078, 3676, 4715, 13, 2223, 327, 1355, 15302, 273, 247, 1643, 4289, 3530, 12262, 247, 1798, 2898, 27, 12805, 356, 965, 6836, 1162, 355, 15, 544, 1047, 62, 3559, 247, 29072, 14, 3169, 2746, ...
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true...
[ 0, 3, 18, 21, 28, 33, 37, 42, 50, 57, 62, 71, 72, 78, 81, 87, 96, 99, 101, 105, 113, 121, 131, 133, 144, 156, 157, 162, 164, 166, 169, 172, 175, 176, 178, 180, 181, 191, 193, 201, 202, 207, 216, 222, 224, 232, 235, 242, ...
[ 3, 18, 21, 28, 33, 37, 42, 50, 57, 62, 71, 72, 78, 81, 87, 96, 99, 101, 105, 113, 121, 131, 133, 144, 156, 157, 162, 164, 166, 169, 172, 175, 176, 178, 180, 181, 191, 193, 201, 202, 207, 216, 222, 224, 232, 235, 242, 246, ...
[ "The", " interpretation", " of", " stroke", " data", " has", " been", " pursued", " before", " deep", " learning", ",", " often", " on", " small", " datasets", " of", " a", " few", " hundred", " samples", " targeting", " a", " particular", " application", ":", " C...
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
[ "The", "interpretation", "of", "stroke", "data", "has", "been", "pursued", "before", "deep", "learning,", "often", "on", "small", "datasets", "of", "a", "few", "hundred", "samples", "targeting", "a", "particular", "application:", "Costagliola", "et", "al.", "[14...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ 0, 4, 19, 22, 29, 34, 38, 43, 51, 58, 63, 73, 79, 82, 88, 97, 100, 102, 106, 114, 122, 132, 134, 145, 158, 170, 173, 177, 182, 192, 194, 208, 217, 223, 225, 233, 236, 243, 247, 255, 261, 268, 272, 280, 284, 298, 309, 313, ...
[ 3, 18, 21, 28, 33, 37, 42, 50, 57, 62, 72, 78, 81, 87, 96, 99, 101, 105, 113, 121, 131, 133, 144, 157, 169, 172, 176, 181, 191, 193, 207, 216, 222, 224, 232, 235, 242, 246, 254, 260, 267, 271, 279, 283, 297, 308, 312, 316,...
[ { "start": 0, "end": 379, "text": "The interpretation of stroke data has been pursued before deep learning, often on small datasets of a few hundred samples targeting a particular application: Costagliola et al. [14] presented a parsing-based approach using a grammar of shapes and symbols where shapes a...
723e8f97fde15f7a8d5ff8d558ea3f16/g03/paragraph_generate
723e8f97fde15f7a8d5ff8d558ea3f16/g03
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
1paragraph_generated
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
2
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
gen-1790736978-w3nzAmOf5F2p0xzyhUqf
openai/gpt-6-luna
openai/gpt-6-luna-20260922
default
fully_faithful
fully_faithful
{"citations_preserved":true,"clarity":"better","content_covered":true,"context_fit":"good","issues":[],"meaning_preserved":true,"outline_faithful":true,"source_issues":[]}
{"differences":[],"dimensions":{"attribution_citations":"preserved","explicit_content":"preserved","implications":"preserved","qualifications":"preserved","quantitative_details":"preserved","tone_rhetorical_function":"preserved"},"explanation":"The candidate preserves the claims, examples, scope, number of diagrams, an...
1
The interpretation of stroke data has been pursued before deep learning, often on small datasets of a few hundred samples targeting a particular application: Costagliola et al. [14] presented a parsing-based approach using a grammar of shapes and symbols where shapes and symbols are independently recognized and the res...
e47c1800428573f77a9ff9d1762ac1dad20160c114e38e798f91eded6fd10cb3
8c38dcf2b223346655eb32b595cb9bd804dde2f234aed9cd45385f49dd2285ab
2mixed
true
548
1,172
1ai
unicode_code_points
half_open
[ { "start": 0, "end": 548, "label": 0, "source_label": "human_preserved", "source_start": 0, "source_end": 548 }, { "start": 548, "end": 1172, "label": 1, "source_label": "ai_rewritten", "source_start": 548, "source_end": 1115 }, { "start": 1172, "end":...
[ 510, 7914, 273, 9980, 941, 556, 644, 23321, 1078, 3676, 4715, 13, 2223, 327, 1355, 15302, 273, 247, 1643, 4289, 3530, 12262, 247, 1798, 2898, 27, 12805, 356, 965, 6836, 1162, 355, 15, 544, 1047, 62, 3559, 247, 29072, 14, 3169, 2746, ...
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true...
[ 0, 3, 18, 21, 28, 33, 37, 42, 50, 57, 62, 71, 72, 78, 81, 87, 96, 99, 101, 105, 113, 121, 131, 133, 144, 156, 157, 162, 164, 166, 169, 172, 175, 176, 178, 180, 181, 191, 193, 201, 202, 207, 216, 222, 224, 232, 235, 242, ...
[ 3, 18, 21, 28, 33, 37, 42, 50, 57, 62, 71, 72, 78, 81, 87, 96, 99, 101, 105, 113, 121, 131, 133, 144, 156, 157, 162, 164, 166, 169, 172, 175, 176, 178, 180, 181, 191, 193, 201, 202, 207, 216, 222, 224, 232, 235, 242, 246, ...
[ "The", " interpretation", " of", " stroke", " data", " has", " been", " pursued", " before", " deep", " learning", ",", " often", " on", " small", " datasets", " of", " a", " few", " hundred", " samples", " targeting", " a", " particular", " application", ":", " C...
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
[ "The", "interpretation", "of", "stroke", "data", "has", "been", "pursued", "before", "deep", "learning,", "often", "on", "small", "datasets", "of", "a", "few", "hundred", "samples", "targeting", "a", "particular", "application:", "Costagliola", "et", "al.", "[14...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ 0, 4, 19, 22, 29, 34, 38, 43, 51, 58, 63, 73, 79, 82, 88, 97, 100, 102, 106, 114, 122, 132, 134, 145, 158, 170, 173, 177, 182, 192, 194, 208, 217, 223, 225, 233, 236, 243, 247, 255, 261, 268, 272, 280, 284, 298, 309, 313, ...
[ 3, 18, 21, 28, 33, 37, 42, 50, 57, 62, 72, 78, 81, 87, 96, 99, 101, 105, 113, 121, 131, 133, 144, 157, 169, 172, 176, 181, 191, 193, 207, 216, 222, 224, 232, 235, 242, 246, 254, 260, 267, 271, 279, 283, 297, 308, 312, 316,...
[ { "start": 0, "end": 379, "text": "The interpretation of stroke data has been pursued before deep learning, often on small datasets of a few hundred samples targeting a particular application: Costagliola et al. [14] presented a parsing-based approach using a grammar of shapes and symbols where shapes a...
723e8f97fde15f7a8d5ff8d558ea3f16/g04/human_original
723e8f97fde15f7a8d5ff8d558ea3f16/g04
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
0human_original
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
9
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
false
needs_researcher_review
[ "audit_source_extraction_issue" ]
null
null
null
null
not_applicable
fully_faithful
{"citations_preserved":true,"clarity":"better","content_covered":true,"context_fit":"good","issues":[],"meaning_preserved":true,"outline_faithful":true,"source_issues":["The preceding paragraph contains obvious PDF extraction corruption: the Fig. 9 caption is interleaved with surrounding text, and symbols and words are...
{"differences":[],"dimensions":{"attribution_citations":"preserved","explicit_content":"preserved","implications":"preserved","qualifications":"preserved","quantitative_details":"preserved","tone_rhetorical_function":"preserved"},"explanation":"The candidate preserves the comparison with the baselines, the specific Ske...
1
Number of GMM components. Using a multi-modal distribution to model the stroke embedding predictions significantly improves our model’s performance (cf. Fig. 9). It is an important hyper-parameter as it is the only source of stochasticity in our relational model R θ . We observe that using 10 or more components is suff...
ad6bcae36b2fe3845ee0f3219ab8cdce0830bec0abfd3c68d9613eb5866948ea
ad6bcae36b2fe3845ee0f3219ab8cdce0830bec0abfd3c68d9613eb5866948ea
0human
false
1,097
1,922
0human
unicode_code_points
half_open
[ { "start": 0, "end": 3022, "label": 0, "source_label": "human_preserved", "source_start": 0, "source_end": 3022 } ]
[ 8447, 273, 443, 8643, 4295, 15, 6915, 247, 4471, 14, 24353, 3268, 281, 1566, 253, 9980, 21496, 13650, 3012, 19132, 776, 1566, 457, 84, 3045, 313, 7836, 15, 2667, 15, 898, 481, 733, 310, 271, 1774, 4373, 14, 19484, 347, 352, 310, 253...
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true...
[ 0, 6, 9, 11, 13, 24, 25, 31, 33, 39, 40, 45, 58, 61, 67, 71, 78, 88, 100, 114, 123, 127, 133, 134, 135, 147, 149, 151, 152, 156, 157, 159, 161, 164, 167, 170, 180, 186, 187, 196, 199, 202, 205, 209, 214, 221, 224, 235, 2...
[ 6, 9, 11, 13, 24, 25, 31, 33, 39, 40, 45, 58, 61, 67, 71, 78, 88, 100, 114, 123, 127, 133, 134, 135, 147, 149, 151, 152, 156, 157, 159, 161, 164, 167, 170, 180, 186, 187, 196, 199, 202, 205, 209, 214, 221, 224, 235, 238, ...
[ "Number", " of", " G", "MM", " components", ".", " Using", " a", " multi", "-", "modal", " distribution", " to", " model", " the", " stroke", " embedding", " predictions", " significantly", " improves", " our", " model", "’", "s", " performance", " (", "cf", "."...
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
[ "Number", "of", "GMM", "components.", "Using", "a", "multi-modal", "distribution", "to", "model", "the", "stroke", "embedding", "predictions", "significantly", "improves", "our", "model’s", "performance", "(cf.", "Fig.", "9).", "It", "is", "an", "important", "hype...
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
[ 0, 7, 10, 14, 26, 32, 34, 46, 59, 62, 68, 72, 79, 89, 101, 115, 124, 128, 136, 148, 153, 158, 162, 165, 168, 171, 181, 197, 200, 203, 206, 210, 215, 222, 225, 239, 242, 246, 257, 263, 265, 267, 269, 272, 280, 285, 291, 294...
[ 6, 9, 13, 25, 31, 33, 45, 58, 61, 67, 71, 78, 88, 100, 114, 123, 127, 135, 147, 152, 157, 161, 164, 167, 170, 180, 196, 199, 202, 205, 209, 214, 221, 224, 238, 241, 245, 256, 262, 264, 266, 268, 271, 279, 284, 290, 293, 29...
[ { "start": 0, "end": 25, "text": "Number of GMM components.", "label": 0, "binary_label": 0, "loss_mask": true, "ai_character_fraction": 0, "human_characters": 25, "ai_characters": 0 }, { "start": 26, "end": 152, "text": "Using a multi-modal distribution to model ...
723e8f97fde15f7a8d5ff8d558ea3f16/g04/paragraph_generate
723e8f97fde15f7a8d5ff8d558ea3f16/g04
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
1paragraph_generated
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
9
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
false
needs_researcher_review
[ "audit_source_extraction_issue" ]
gen-1790736978-Y4DEhm0pnOzaiULuJ2Ap
openai/gpt-6-luna
openai/gpt-6-luna-20260922
default
fully_faithful
fully_faithful
"{\"citations_preserved\":true,\"clarity\":\"better\",\"content_covered\":true,\"context_fit\":\"goo(...TRUNCATED)
"{\"differences\":[],\"dimensions\":{\"attribution_citations\":\"preserved\",\"explicit_content\":\"(...TRUNCATED)
1
"Number of GMM components. Using a multi-modal distribution to model the stroke embedding prediction(...TRUNCATED)
2b20b80fccab9a8dbf3890027e6f9186829218916e16ec4f46fed59182eb06ff
ad6bcae36b2fe3845ee0f3219ab8cdce0830bec0abfd3c68d9613eb5866948ea
2mixed
true
1,097
1,844
1ai
unicode_code_points
half_open
[{"start":0,"end":1097,"label":0,"source_label":"human_preserved","source_start":0,"source_end":1097(...TRUNCATED)
[8447,273,443,8643,4295,15,6915,247,4471,14,24353,3268,281,1566,253,9980,21496,13650,3012,19132,776,(...TRUNCATED)
[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true(...TRUNCATED)
[0,6,9,11,13,24,25,31,33,39,40,45,58,61,67,71,78,88,100,114,123,127,133,134,135,147,149,151,152,156,(...TRUNCATED)
[6,9,11,13,24,25,31,33,39,40,45,58,61,67,71,78,88,100,114,123,127,133,134,135,147,149,151,152,156,15(...TRUNCATED)
["Number"," of"," G","MM"," components","."," Using"," a"," multi","-","modal"," distribution"," to"(...TRUNCATED)
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
["Number","of","GMM","components.","Using","a","multi-modal","distribution","to","model","the","stro(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[0,7,10,14,26,32,34,46,59,62,68,72,79,89,101,115,124,128,136,148,153,158,162,165,168,171,181,197,200(...TRUNCATED)
[6,9,13,25,31,33,45,58,61,67,71,78,88,100,114,123,127,135,147,152,157,161,164,167,170,180,196,199,20(...TRUNCATED)
[{"start":0,"end":25,"text":"Number of GMM components.","label":0,"binary_label":0,"loss_mask":true,(...TRUNCATED)
723e8f97fde15f7a8d5ff8d558ea3f16/g05/human_original
723e8f97fde15f7a8d5ff8d558ea3f16/g05
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
0human_original
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
1
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
null
null
null
null
not_applicable
fully_faithful
"{\"citations_preserved\":true,\"clarity\":\"better\",\"content_covered\":true,\"context_fit\":\"goo(...TRUNCATED)
"{\"differences\":[],\"dimensions\":{\"attribution_citations\":\"not_applicable\",\"explicit_content(...TRUNCATED)
1
"We present a generative model for complex free-form structures such as stroke-based drawing tasks. (...TRUNCATED)
86365deec75ef64aa4bb22c314cd06d1e1f48aaddbc930c2f3bb4f5bced27cc4
86365deec75ef64aa4bb22c314cd06d1e1f48aaddbc930c2f3bb4f5bced27cc4
0human
false
1,015
1,583
0human
unicode_code_points
half_open
[{"start":0,"end":2105,"label":0,"source_label":"human_preserved","source_start":0,"source_end":2105(...TRUNCATED)
[1231,1246,247,1006,800,1566,323,2570,1959,14,630,5289,824,347,9980,14,3169,10263,8892,15,3900,2045,(...TRUNCATED)
[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true(...TRUNCATED)
[0,2,10,12,18,23,29,33,41,46,47,51,62,67,70,77,78,83,91,97,98,104,113,124,129,132,141,142,147,154,15(...TRUNCATED)
[2,10,12,18,23,29,33,41,46,47,51,62,67,70,77,78,83,91,97,98,104,113,124,129,132,141,142,147,154,158,(...TRUNCATED)
["We"," present"," a"," gener","ative"," model"," for"," complex"," free","-","form"," structures","(...TRUNCATED)
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
["We","present","a","generative","model","for","complex","free-form","structures","such","as","strok(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[0,3,11,13,24,30,34,42,52,63,68,71,84,92,99,105,114,125,130,133,148,155,159,168,171,177,185,188,200,(...TRUNCATED)
[2,10,12,23,29,33,41,51,62,67,70,83,91,98,104,113,124,129,132,147,154,158,167,170,176,184,187,199,20(...TRUNCATED)
[{"start":0,"end":98,"text":"We present a generative model for complex free-form structures such as (...TRUNCATED)
723e8f97fde15f7a8d5ff8d558ea3f16/g05/paragraph_generate
723e8f97fde15f7a8d5ff8d558ea3f16/g05
723e8f97fde15f7a8d5ff8d558ea3f16
null
train
1paragraph_generated
CoSE: Compositional Stroke Embeddings
[ "Aksan, Emre", "Deselaers, Thomas", "Tagliasacchi, Andrea", "Hilliges, Otmar" ]
NeurIPS
2,020
1
https://proceedings.neurips.cc/paper_files/paper/2020/file/723e8f97fde15f7a8d5ff8d558ea3f16-Paper.pdf
https://proceedings.neurips.cc/paper_files/paper/2020/hash/723e8f97fde15f7a8d5ff8d558ea3f16-Abstract.html
e39d62581c9cce0f20a188b8fe1960a0463cbb552da7a39c963b833b52e01198
true
true
automated_checks_passed_pending_expert_review
[]
gen-1790736978-A8i5B2fLY9WUjz1QiHoJ
openai/gpt-6-luna
openai/gpt-6-luna-20260922
default
fully_faithful
fully_faithful
"{\"citations_preserved\":true,\"clarity\":\"better\",\"content_covered\":true,\"context_fit\":\"goo(...TRUNCATED)
"{\"differences\":[],\"dimensions\":{\"attribution_citations\":\"not_applicable\",\"explicit_content(...TRUNCATED)
1
"We present a generative model for complex free-form structures such as stroke-based drawing tasks. (...TRUNCATED)
d4a1424e67a15e388a01ec765abe5ed261ef7a4e2cb51eeaf0d01e878736d577
86365deec75ef64aa4bb22c314cd06d1e1f48aaddbc930c2f3bb4f5bced27cc4
2mixed
true
1,015
1,625
1ai
unicode_code_points
half_open
[{"start":0,"end":1015,"label":0,"source_label":"human_preserved","source_start":0,"source_end":1015(...TRUNCATED)
[1231,1246,247,1006,800,1566,323,2570,1959,14,630,5289,824,347,9980,14,3169,10263,8892,15,3900,2045,(...TRUNCATED)
[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true,true(...TRUNCATED)
[0,2,10,12,18,23,29,33,41,46,47,51,62,67,70,77,78,83,91,97,98,104,113,124,129,132,141,142,147,154,15(...TRUNCATED)
[2,10,12,18,23,29,33,41,46,47,51,62,67,70,77,78,83,91,97,98,104,113,124,129,132,141,142,147,154,158,(...TRUNCATED)
["We"," present"," a"," gener","ative"," model"," for"," complex"," free","-","form"," structures","(...TRUNCATED)
9fd55248d51d33976b324fc11592e28071da7d41e0e9401dfb7082e30574b7b1
["We","present","a","generative","model","for","complex","free-form","structures","such","as","strok(...TRUNCATED)
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)
[0,3,11,13,24,30,34,42,52,63,68,71,84,92,99,105,114,125,130,133,148,155,159,168,171,177,185,188,200,(...TRUNCATED)
[2,10,12,23,29,33,41,51,62,67,70,83,91,98,104,113,124,129,132,147,154,158,167,170,176,184,187,199,20(...TRUNCATED)
[{"start":0,"end":98,"text":"We present a generative model for complex free-form structures such as (...TRUNCATED)
End of preview. Expand in Data Studio

AI Research Paper Provenance: v3

2,000 papers; 10,000 generated paragraphs + 10,000 matched human originals, for contextual token/sentence detection of LLM-produced wording. Each passage contains the target and its unchanged human neighbors. Flagged examples are retained.

Load and choose a view

from datasets import load_dataset
repo = "woog/ai-paper-provenance-v3"
data = load_dataset(repo, "fresh_passages")

The repository is public; no authentication is required.

View Train Validation Test
passages 12,000 4,000 4,000
fresh_passages 11,700 3,900 3,900
paragraphs 12,000 4,000 4,000
sentences 159,974 53,599 53,223

fresh_passages is the recommended context-preserving view: it excludes development-exposed pilots. paragraphs and sentences are target-only and sentence views; filter development_exposed == false yourself. These are alternate views of the same examples, not independent datasets.

Labels and identifiers

  • Token loss: token_labels uses human 0, AI 1, ignored -100; token_loss_mask identifies supervised tokens. Special/padding tokens must also receive -100.
  • Three-way provenance: human 0, AI 1, mixed 2. Passage label=2 means human context plus an AI target; it is not a token target. Use target_label for the target paragraph and has_ai for binary passage detection.
  • Mixed boundaries: 79 sentences are mixed; their binary label is -100. There are 3 ignored tokens. Preserve these masks.
  • Provenance rule: the entire generated target is AI-labeled, including wording copied from its source. Unchanged surrounding paragraphs are human. Production history is not always inferable from text.
  • Alignment: input_ids, token labels/masks, and token start/end offsets align with the bundled tokenizer, without special tokens, padding, or truncation. For another tokenizer, rebuild labels from character regions; do not reuse these IDs. Offsets are half-open Unicode code points [start,end), not bytes or UTF-16 positions.
  • IDs: paper_id is the stable corpus ID; pair_id joins matched variants. Nullable forum_id is the verified scraped OpenReview ID; null means no verified match. See identifier_mapping.json.

Recommended training and evaluation recipe

This is a starting protocol, not a claim that hyperparameters or mixture weights have been optimized.

  1. Filter whole pairs. Start with pairs where both variants pass eligible_for_pilot_training and paired_generation_fidelity_verdict is fully_faithful or mostly_faithful_with_minor_differences. Use that paired field for human rows too. Report retained counts, a fully-faithful-only sensitivity check, and broader flagged-data results separately.
  2. Use text-only inputs and masked token cross-entropy. Feed one variant at a time. Originals, notes, judgments and metadata are not classifier inputs. Pairing controls topic; it does not prescribe contrastive loss. Sentence-only classification is a separate baseline.
  3. Sample by paper/target; balance variants. Mix target-only examples with varied-offset context windows, covering boundaries and human-only text. Preserve offsets/masks and avoid truncating away targets. Token classes remain imbalanced; tune weighting on validation data. This reduces the fixed-middle-position shortcut.
  4. Control duplicates. For paragraph/sentence baselines, exclude text_has_multiple_labels from the primary subset. Deduplicate or weight repeated context using training-only frequencies. For strict text-disjoint evaluation, exclude text_appears_in_multiple_splits; exported duplicate weights describe the whole view.
  5. Calibrate on validation, then freeze. Choose thresholds and sentence aggregation/smoothing at a stated human false-positive target. Report token/sentence precision, recall, false positives and boundary performance, with uncertainty clustered by paper. Preserve mixed-boundary masks.
  6. Keep transfer tests and mining separate. Train/tune on Luna; test expensive generators on unseen held-out papers. Mine human false positives only from a separate training/development pool, add synthetic mirrors, and retrain. Never mine the final test set.

How this differs from Pangram 4

Our labels describe binary surface-production provenance. Pangram distinguishes human, AI-assisted and AI-generated text. Our mixed=2 must not map to AI-assisted: it describes composition/boundaries. Our reconstructions preserve human-specified ideas and could qualify as assisted under Pangram's definition. Reproducing its taxonomy requires additional labeling and data.

Aspect This dataset Pangram 4 reference
Generation Luna reconstructs a hidden paragraph from detailed notes, abstract and neighbors Broader generator/domain mixture, synthetic mirrors and edited text
Labels Known replacement spans; copied output remains AI-labeled Clause-level lexical/semantic matching distinguishes human, assisted and generated text
Training Supplies supervision; no required backbone or tuned schedule Segment-first training, then tokenwise and auxiliary objectives
Causal context Choose a suitable contextual classifier Repeat2 duplicates each sampled window and masks loss on the first copy

Pangram uses 512-token source windows and inference stride 256. Repeat2 supervises only the second copy; a bidirectional encoder does not require it. This corpus alone cannot reproduce its humanizer or full three-class objectives. See Pangram 4 §§2–4.

Splits, quality and limitations

Paper splits are 1,200 train / 400 validation / 400 test. All five targets, variants and derived rows stay with their paper; normalized author identities do not cross splits. The newest 1,500 papers comprise 500 each from ACL, ICML and NeurIPS, 2013–2021. The original 250 pilot targets are marked development_exposed=true; fresh_passages excludes their 500 rows. See provenance.json for sampling and hashes.

The existing eligibility flag passes 6,375 passage rows, including 2,739 generated paragraphs, before the paired fidelity filter above. Flags apply to parent passages, not individual sentence defects. Luna reviews its own model family; these are not expert gold labels. Historical dates support human provenance but do not prove it. PDF extraction can introduce artifacts.

The writer sees the abstract, immediate neighbors, structured notes, and a broad length range, but never the held-out paragraph directly. A separate call extracts notes from that paragraph. First mechanically valid outputs are retained; there are no quality-based rerolls. Full prompts are in generation_protocol.json.

The full view contains 1,273,649 AI-labeled and 6,789,351 human-labeled tokens. The latest expansion cost $7.1463, using 50,791,420 input / 15,984,194 output API tokens, including notes, retries and reviews. All accepted new outputs were verified as Flex. See costs.json and statistics.json for details. Source/model/tokenizer rights are described in LICENSE.md; this corpus does not establish generalization to other generators or writing workflows.

Downloads last month
100