Datasets:
id string | document_id string | images images list | question string | answer string | answer_format string | conversations string | kind string | source_name string | answerable bool | evidence_pages list | related_context_pages list | token_count int64 | page_count int64 | source_domain string | domain_tag_method string | document_language string | source_title string | source_authors list | source_url string | license string | license_url string | context_sha256 string | model string | generation_method string | verification_policy string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1703.00544v5_reasoning_000 | 1703.00544v5 | In the contribution overview of Section 1.2, how many numbered named results are stated from Theorem 1.1 through Theorem 1.7, counting both the theorem and corollary statements shown there? | 7 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
2,
3
] | [] | 81,132 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_001 | 1703.00544v5 | In Table 1, how many green FPT cells are shown across both parameter blocks? | 5 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
4
] | [] | 81,105 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_002 | 1703.00544v5 | In Section 2.1, before the pre-evaluations paragraph, how many cardinality-constraint categories are introduced by the consecutive headings “Global cardinality constraints” and “Local cardinality constraints”? | 2 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
5,
6
] | [] | 81,128 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_003 | 1703.00544v5 | According to Definition 2.4, how many node types are listed for a nice tree decomposition? | 4 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
7
] | [] | 81,107 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_004 | 1703.00544v5 | In the refinement discussion, multiply the number of cases in Definition 3.5’s shape definition by the maximum number of parts into which Lemma 3.8 can split a nonuniform type. What is the product? | 8 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
13,
14
] | [] | 81,131 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_005 | 1703.00544v5 | In the definition of a parameterized reduction in Section 2.3, how many bullet-point conditions are listed after “such that”? | 2 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
8
] | [] | 81,114 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_006 | 1703.00544v5 | In the proof of Theorem 3.1, add the number of input items named in the LCC Subset problem box to the number of vertex groups listed in the construction of the graph H. What is the sum? | 6 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
9,
10
] | [] | 81,132 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_007 | 1703.00544v5 | In the displayed definition of the formula φ at the end of the proof of Theorem 1.3, how many conditions are conjoined for each v in C? | 2 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
12
] | [] | 81,121 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_008 | 1703.00544v5 | Combining Proposition 3.7 and Lemma 3.10, how many bullet-point conclusions are listed in total? | 4 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
14,
15
] | [] | 81,112 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_009 | 1703.00544v5 | In Theorem 3.11’s integer linear program description, how many displayed constraint labels are shown from (0) through the shape constraints? | 5 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
16
] | [] | 81,117 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_010 | 1703.00544v5 | How many more component bullets are listed in Definition 4.1 for a CSP instance than output-property bullets are listed in Theorem 4.6? | 1 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
21,
22
] | [] | 81,118 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1703.00544v5_reasoning_011 | 1703.00544v5 | How many labeled subfigures does Figure 3 contain? | 2 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
28
] | [] | 81,098 | 32 | Computer Science | majority of full-context generator domain labels | en | Simplified Algorithmic Metatheorems Beyond MSO: Treewidth and Neighborhood Diversity | [
"Du\\v{s}an Knop, Martin Kouteck\\'y, Tom\\'a\\v{s} Masa\\v{r}\\'ik, Tom\\'a\\v{s} Toufar"
] | https://arxiv.org/abs/1703.00544v5 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6c3998e25a5a47a1839c8784d1b1da7ed3417fd195730e4623413eeea6b7c35e | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_000 | 1906.02754v2 | In the final-state object definitions, what is the numerical difference between the global Higgs rapidity cutoff |Y_H| < 2.5 and the truth-jet radius R = 0.4? | 2.1 | Float | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
4,
5
] | [] | 36,563 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_001 | 1906.02754v2 | On the title page author-affiliation list, how many numbered affiliations are located in either France or the USA? | 5 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
0
] | [] | 36,542 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_002 | 1906.02754v2 | Using the prose bin definitions for gg → H and electroweak qqH production, what is the difference in GeV between the largest m_jj boundary listed for the electroweak qqH m_jj > 350 GeV bin and the gg → H low/high-m_jj separation? | 1150 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
6,
8
] | [] | 36,586 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_003 | 1906.02754v2 | In the Contents, how many entries are listed under “4 Bin Definitions” when counting both its direct subsections and the two indented subsubsections under “Other Production Modes”? | 7 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
1
] | [] | 36,555 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_004 | 1906.02754v2 | In the stage 1.2 diagrams, what is the combined number of boxes immediately below the top process box for electroweak qqH and leptonic VH production? | 6 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
7,
9
] | [] | 36,554 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_005 | 1906.02754v2 | In the Overview’s list of concrete STXS-bin goals, how many goals are given? | 4 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
3
] | [] | 36,537 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_006 | 1906.02754v2 | In the final-state object definitions, what is the numerical difference between the forward-electron rapidity reach mentioned for H → ZZ* → 4ℓ and the truth-jet radius R? | 4.6 | Float | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
4,
5
] | [] | 36,559 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_007 | 1906.02754v2 | In Figure 1 for gg → H, how many top-level p_T^H ranges split the stage 1.2 diagram before the jet-category branches? | 2 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
5
] | [] | 36,551 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_008 | 1906.02754v2 | What is the difference in GeV between the highest finite p_T^H boundary shown for ttH production and the highest p_T^V boundary stated for the q q′ → WH and q q̄ → ZH subprocesses in associated leptonic VH production? | 50 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
9,
10
] | [] | 36,574 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_009 | 1906.02754v2 | In the electroweak qqH production description, into how many jet bins is the cross section first split? | 3 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
7
] | [] | 36,542 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_010 | 1906.02754v2 | If each dominant Higgs production process named in the abstract were paired with each major Higgs decay channel listed in the Introduction, how many production–decay pairings would that make? | 20 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
0,
1
] | [] | 36,556 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
1906.02754v2_reasoning_011 | 1906.02754v2 | In Figure 4 for ttH production, how many finite p_T^H boundary labels greater than 0 are shown? | 5 | Integer | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | reasoning | mmprolong_public_reasoning | true | [
10
] | [] | 36,544 | 14 | Physics | majority of full-context generator domain labels | en | Simplified Template Cross Sections -- Stage 1.1 and 1.2 | [
"Nicolas Berger, Claudia Bertella, Matteo Bonanomi, Nihal Brahimi, Thomas P. Calvet, Milene Calvetti, Valerio Dao, Marco Delmastro, Michael Duehrssen-Debling, Paolo Francavilla, Yacine Haddad, Sarah Heim, Jelena Jovicevic, Oleh Kivernyk, Maria Moreno Llacer, Jonathon M. Langford, Changqiao Li, Giovanni Marchiori, J... | https://arxiv.org/abs/1906.02754v2 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | b128c9c48a5281953ecbec01acc88407c12dcfff2fe9a0de99c363bcc6c79fb6 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
2003.13227v3_multi_distributed_006 | 2003.13227v3 | Compare Theorem 1.1 with the single-closed-subset version Theorem 3.4: what restriction and distance requirements do they impose on the constructed metric? | ["Theorem 1.1: the constructed metric m satisfies m|A_i^2 = e_i for every i in I, and D_X(m,d) = sup_i D_{A_i}(e_i,d|A_i^2).", "Theorem 3.4: the constructed metric m satisfies m|A^2 = e, and D_X(m,d) = D_A(e,d|A^2)."] | List | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | multi_distributed | mmprolong_public_multi_distributed | true | [
1,
7
] | [] | 57,474 | 22 | Mathematics | majority of full-context generator domain labels | en | An interpolation of metrics and spaces of metrics | [
"Yoshito Ishiki"
] | https://arxiv.org/abs/2003.13227v3 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6d7a27fb7247598661b042abf797d7823b23f1dbee7dac91ca11936d341ff5d9 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
2003.13227v3_multi_distributed_011 | 2003.13227v3 | What formulas does the paper give for D_X on Met(X) and for the metric E on C(X^2)? | ["D_X(d,e) = sup_{(x,y) in X^2} |d(x,y) - e(x,y)|.", "E(f,g) = min{1, sup_{(x,y) in X^2} |f(x,y) - g(x,y)|}."] | List | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | multi_distributed | mmprolong_public_multi_distributed | true | [
0,
18
] | [] | 57,435 | 22 | Mathematics | majority of full-context generator domain labels | en | An interpolation of metrics and spaces of metrics | [
"Yoshito Ishiki"
] | https://arxiv.org/abs/2003.13227v3 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6d7a27fb7247598661b042abf797d7823b23f1dbee7dac91ca11936d341ff5d9 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review | |
2003.13227v3_multi_local_010 | 2003.13227v3 | Across the later sections, what conclusions are stated in Theorem 4.7, Theorem 4.15, and Lemma 5.1 about spaces or sets of metrics? | ["Theorem 4.7: for every non-discrete completely metrizable space X, the set of all d in Comp(X) for which (X,d) satisfies the anti-G-transmissible property is dense G_delta in Comp(X).", "Theorem 4.15: for every non-discrete metrizable space X, the set of all metrics d in Met(X) for which (X,d) has rich pseudo-cones i... | List | [{"from": "human", "value": "You are given a document with text and images, and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write 'Not answerable.' Write your answer in the following f... | multi_local | mmprolong_public_multi_local | true | [
11,
15,
18
] | [] | 57,524 | 22 | Mathematics | majority of full-context generator domain labels | en | An interpolation of metrics and spaces of metrics | [
"Yoshito Ishiki"
] | https://arxiv.org/abs/2003.13227v3 | CC-BY-4.0 | http://creativecommons.org/licenses/by/4.0/ | 6d7a27fb7247598661b042abf797d7823b23f1dbee7dac91ca11936d341ff5d9 | gpt-5.5-2026-04-24 | historical_full_document | historical_independent_full_document_review |
MMProLong-Data-Public
This is a separately constructed public corpus, not the original data used in the MMProLong paper. The original corpus cannot be released under ByteDance's internal data governance and management requirements. This public release does not establish reproduction of the paper's reported results.
New production uses GPT-5.6 (gpt-5.6-sol). Historical GPT-5.5 examples retain their original model labels. Please cite the MMProLong paper when using this dataset or its synthesis methodology.
Dataset format
The release consists of this card and data/train-*.parquet. Shards are approximately 5 GB each; the final shard may be smaller. Each row is one QA example and contains every page image of its source PDF, in order. Images are embedded in the parquet as Hugging Face Image values; no separate PDF/image archive or restoration step is required.
The page images use a smaller lossless encoding where beneficial. Every conversion is decoded and checked for identical RGBA pixels and dimensions. All questions, answers, page indices, full-document coverage and stored token counts are preserved. QA from the same document share a parquet image dictionary to reduce physical duplication.
Load the data
from datasets import load_dataset
dataset = load_dataset(
"ZhaoweiWang/MMProLong-Data-Public", split="train", streaming=True, batch_size=1
)
example = next(iter(dataset))
print(example["question"], example["answer"])
print(example["kind"], example["evidence_pages"])
pages = example["images"] # ordered PIL images for the complete PDF
For image bytes instead of decoded images, cast images to datasets.Sequence(datasets.Image(decode=False)). conversations contains the original serialized training dialogue, including the document's image placeholders.
Published statistics
Snapshot date: 2026-11-13. These are measured counts of the published rows.
| Statistic | Value |
|---|---|
| QA examples | 90,556 |
| Source documents | 37,088 |
| Unique PDF pages | 774,132 |
| Estimated training tokens | 4,999,938,857 (5.000B) |
| Average tokens per QA | 55,214 |
| Median tokens per QA | 50,722 |
| Token range per QA | 33,547–125,503 |
| Average pages per document | 20.9 |
| PDF page range | 14–50 |
| Parquet shards | 22 |
Training tokens are the stored Qwen2.5-VL serialized-sample estimate, including page-image context and QA/prompt text. They are not billed API tokens or unique document-text tokens. A document used by several QA examples contributes its full context to each training example.
Tasks
| Task | QA | Mean pages per QA | Mean tokens per QA |
|---|---|---|---|
| Single-page extraction | 32,998 | 21.5 | 54,574 |
| Local multi-page extraction | 15,746 | 22.5 | 57,079 |
| Distributed multi-page extraction | 16,428 | 21.6 | 54,778 |
| Reasoning | 16,369 | 21.7 | 55,005 |
| Unanswerable | 9,015 | 21.9 | 55,469 |
single is single-page extraction; multi_local requires nearby pages; multi_distributed requires separately necessary distant pages; reasoning requires a derived numerical answer; unanswerable requires missing information that cannot be recovered from the complete document. All evidence indices are zero-based PDF page indices. Unanswerable examples have empty evidence_pages and separate related_context_pages.
Sources and domains
| Source | Documents |
|---|---|
| arxiv | 4,386 |
| elife | 20 |
| plos | 32,682 |
| Document domain | Documents | Document share | QA |
|---|---|---|---|
| Medicine | 10,578 | 28.52% | 19,169 |
| Biology | 9,805 | 26.44% | 17,284 |
| Social Science | 4,226 | 11.39% | 8,187 |
| Computer Science | 2,752 | 7.42% | 11,850 |
| Engineering | 2,662 | 7.18% | 6,845 |
| Economics | 2,611 | 7.04% | 5,251 |
| Mathematics | 1,605 | 4.33% | 9,516 |
| Physics | 1,473 | 3.97% | 9,561 |
| Earth Science | 1,236 | 3.33% | 2,576 |
| Language and Literature | 125 | 0.34% | 249 |
| Other | 15 | 0.04% | 68 |
Domain labels describe the source document and are automatic annotations, not human gold labels. domain_tag_method preserves their provenance. The source pool contains public research papers from multiple fields; it is not restricted to computer science. Sampling favors underrepresented domains, but the table reports the actual resulting distribution.
Generation models
| Recorded generation model | QA | Estimated training tokens |
|---|---|---|
| gpt-5.5-2026-04-24 | 28,524 | 1,744,611,453 |
| gpt-5.6-sol | 62,032 | 3,255,327,404 |
Use model to select a model-specific subset. The complete release is a mixed-model historical corpus.
Columns
| Column | Meaning |
|---|---|
id, document_id |
Stable QA and source-document identifiers |
images |
Complete ordered page-image context, embedded bytes with logical filenames |
question, answer, answer_format |
QA target and answer format; structured answers are serialized as JSON text |
conversations |
Original serialized training dialogue |
kind, source_name, answerable |
Task category and answerability |
evidence_pages, related_context_pages |
Necessary evidence or related-context anchors, indexed from zero |
token_count, page_count |
Stored training-token estimate and full PDF page count |
source_domain, domain_tag_method, document_language |
Document annotations and domain-label provenance |
source_title, source_authors, source_url |
Source attribution and retrieval information |
license, license_url, context_sha256 |
Original-document license and identity |
model, generation_method, verification_policy |
Recorded synthesis and review provenance |
Construction and quality checks
Local tasks use the released 8–15-page sampler, original prompts and two LongDocURL few-shot examples. Distributed extraction, globally planned reasoning and unanswerable tasks use a GPT-5.6 document tool loop with overview, full-text search, page-text/image inspection and bounded arithmetic. Unanswerable agents receive every extracted text page and inspect relevant images; a zero-hit literal search is not proof of absence.
New hybrid examples require two independent full-PDF reviews, structural checks, executable arithmetic for numeric reasoning, necessary multi-page evidence and per-document semantic novelty checks. At most two QA are admitted per document in current production. Historical examples preserve their original model, sampling and review policy; some historical documents therefore have more QA.
Automated review is a filter, not a correctness guarantee. Domain/language annotations may contain errors. Benchmark decontamination and a GPU training-read evaluation have not been completed. Evaluate overlap with your own test sets before training.
Source rights
The included source papers were admitted under CC BY 4.0 checks. Per-row title, authors, source URL and license fields preserve attribution. Cite and attribute the underlying documents as required by their licenses, as well as MMProLong. This release contains separately generated public data; it does not distribute the paper's original internal training corpus.
Citation
@inproceedings{wang2026mmprolong,
title={Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context},
author={Wang, Zhaowei and Luo, Lishu and Duan, Haodong and Liu, Weiwei and Wu, Sijin and Luo, Ji and Yan, Shen and Peng, Shuai and Yuan, Sihang and Huang, Chaoyi and Lin, Yi and Song, Yangqiu},
booktitle={Advances in Neural Information Processing Systems (NeurIPS 2026)},
year={2026},
url={https://arxiv.org/abs/2605.13831}
}
- Downloads last month
- 3,251