id stringlengths 20 20 | document_id stringlengths 16 16 | task_type stringclasses 2
values | language stringclasses 1
value | images listlengths 32 50 | question stringlengths 69 287 | answer stringlengths 1 259 | answer_format stringclasses 4
values | segment_pages listlengths 8 15 | evidence listlengths 1 3 | quality dict | messages listlengths 2 2 |
|---|---|---|---|---|---|---|---|---|---|---|---|
9c01d650c9049ffbc463 | 5c7209bf5a684f99 | extract_single | en | [
"images/5c7209bf5a684f99/page_0001.jpg",
"images/5c7209bf5a684f99/page_0002.jpg",
"images/5c7209bf5a684f99/page_0003.jpg",
"images/5c7209bf5a684f99/page_0004.jpg",
"images/5c7209bf5a684f99/page_0005.jpg",
"images/5c7209bf5a684f99/page_0006.jpg",
"images/5c7209bf5a684f99/page_0007.jpg",
"images/5c7209b... | According to the “Transformer Equivalence” section on Page 5, what recognizes a language defined by a TL[#] formula of depth k? | a fixed-precision transformer of depth k | string | [
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14
] | [
{
"page": 5,
"kind": "Text",
"description": "Theorem 3.1 in the “Transformer Equivalence” section states that a language defined by a TL[#] formula of depth k is recognized by a fixed-precision transformer of depth k.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/5c7209bf5a684f99/page_0001.jpg"
},
{
"type": "image",
"image": "images/5c7209bf5a684f99/page_0002.jpg"
},
{
"type": "image",
"image": "images/5c7209bf5a684f99/page... |
3b3e12886a856b246c9a | f07f7cd76603e208 | extract_single | en | [
"images/f07f7cd76603e208/page_0001.jpg",
"images/f07f7cd76603e208/page_0002.jpg",
"images/f07f7cd76603e208/page_0003.jpg",
"images/f07f7cd76603e208/page_0004.jpg",
"images/f07f7cd76603e208/page_0005.jpg",
"images/f07f7cd76603e208/page_0006.jpg",
"images/f07f7cd76603e208/page_0007.jpg",
"images/f07f7cd... | According to the section “A.5. Selection of the Simulation Horizon” on Page 26, what default simulation horizon was used for the experiments, in minutes? | 500 | integer | [
15,
16,
17,
18,
19,
20,
21,
22,
23,
24,
25,
26,
27,
28
] | [
{
"page": 26,
"kind": "Text",
"description": "In the section “A.5. Selection of the Simulation Horizon,” the text states that the default experiments use T_sim = 500 min.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/f07f7cd76603e208/page_0001.jpg"
},
{
"type": "image",
"image": "images/f07f7cd76603e208/page_0002.jpg"
},
{
"type": "image",
"image": "images/f07f7cd76603e208/page... |
a23e4b45021badac886f | 7dcb3ca5c0976849 | extract_single | en | [
"images/7dcb3ca5c0976849/page_0001.jpg",
"images/7dcb3ca5c0976849/page_0002.jpg",
"images/7dcb3ca5c0976849/page_0003.jpg",
"images/7dcb3ca5c0976849/page_0004.jpg",
"images/7dcb3ca5c0976849/page_0005.jpg",
"images/7dcb3ca5c0976849/page_0006.jpg",
"images/7dcb3ca5c0976849/page_0007.jpg",
"images/7dcb3ca... | According to Table 4, “OED Productivity, FY89–FY92,” how many adjustment audits were recorded in FY92? | 28 | integer | [
22,
23,
24,
25,
26,
27,
28,
29,
30,
31,
32,
33,
34
] | [
{
"page": 28,
"kind": "Table",
"description": "In Table 4, “OED Productivity, FY89–FY92,” the Adjustment Audits row shows 28 under FY92.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/7dcb3ca5c0976849/page_0001.jpg"
},
{
"type": "image",
"image": "images/7dcb3ca5c0976849/page_0002.jpg"
},
{
"type": "image",
"image": "images/7dcb3ca5c0976849/page... |
d887f26cafe5a265a2f4 | 6e06f5d9853720a1 | extract_single | en | [
"images/6e06f5d9853720a1/page_0001.jpg",
"images/6e06f5d9853720a1/page_0002.jpg",
"images/6e06f5d9853720a1/page_0003.jpg",
"images/6e06f5d9853720a1/page_0004.jpg",
"images/6e06f5d9853720a1/page_0005.jpg",
"images/6e06f5d9853720a1/page_0006.jpg",
"images/6e06f5d9853720a1/page_0007.jpg",
"images/6e06f5d... | According to Table 5 in the Results and Discussion section on Page 20, what PLCC value did MetaQAP achieve on the BIQ2021 dataset? | 0.884 | float | [
17,
18,
19,
20,
21,
22,
23,
24,
25,
26,
27
] | [
{
"page": 20,
"kind": "Table",
"description": "Table 5 in the Results and Discussion section lists MetaQAP's BIQ2021 PLCC as 0.884.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/6e06f5d9853720a1/page_0001.jpg"
},
{
"type": "image",
"image": "images/6e06f5d9853720a1/page_0002.jpg"
},
{
"type": "image",
"image": "images/6e06f5d9853720a1/page... |
22fabb1e39bcabbe49a6 | db7293ba70ded2f6 | extract_single | en | [
"images/db7293ba70ded2f6/page_0001.jpg",
"images/db7293ba70ded2f6/page_0002.jpg",
"images/db7293ba70ded2f6/page_0003.jpg",
"images/db7293ba70ded2f6/page_0004.jpg",
"images/db7293ba70ded2f6/page_0005.jpg",
"images/db7293ba70ded2f6/page_0006.jpg",
"images/db7293ba70ded2f6/page_0007.jpg",
"images/db7293b... | According to the “Docker sandbox for result checking” section on Page 11, what are the two key functionalities of the sandbox environment? | Executing LLM-generated Python code and returning its output in the form of a materials property dictionary; Executing corresponding verification scripts to assess correctness | list | [
9,
10,
11,
12,
13,
14,
15,
16
] | [
{
"page": 11,
"kind": "Text",
"description": "Under the section “Docker sandbox for result checking,” the text explicitly lists the sandbox’s two key functionalities.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/db7293ba70ded2f6/page_0001.jpg"
},
{
"type": "image",
"image": "images/db7293ba70ded2f6/page_0002.jpg"
},
{
"type": "image",
"image": "images/db7293ba70ded2f6/page... |
2aa08ec942ab930aed02 | 6ee5ee8566ebfc2d | extract_single | en | [
"images/6ee5ee8566ebfc2d/page_0001.jpg",
"images/6ee5ee8566ebfc2d/page_0002.jpg",
"images/6ee5ee8566ebfc2d/page_0003.jpg",
"images/6ee5ee8566ebfc2d/page_0004.jpg",
"images/6ee5ee8566ebfc2d/page_0005.jpg",
"images/6ee5ee8566ebfc2d/page_0006.jpg",
"images/6ee5ee8566ebfc2d/page_0007.jpg",
"images/6ee5ee8... | According to the “Blind evaluation by UF Health Oncologists” section on Page 25, how many treatment plans were included in total? | 237 | integer | [
22,
23,
24,
25,
26,
27,
28,
29,
30,
31,
32,
33
] | [
{
"page": 25,
"kind": "Text",
"description": "Under the “Blind evaluation by UF Health Oncologists” section, the study design states that the mixed treatment plans resulted in a total of 237 plans.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/6ee5ee8566ebfc2d/page_0001.jpg"
},
{
"type": "image",
"image": "images/6ee5ee8566ebfc2d/page_0002.jpg"
},
{
"type": "image",
"image": "images/6ee5ee8566ebfc2d/page... |
40fa130bf3d4ec69dc68 | b28830acee54c369 | extract_single | en | [
"images/b28830acee54c369/page_0001.jpg",
"images/b28830acee54c369/page_0002.jpg",
"images/b28830acee54c369/page_0003.jpg",
"images/b28830acee54c369/page_0004.jpg",
"images/b28830acee54c369/page_0005.jpg",
"images/b28830acee54c369/page_0006.jpg",
"images/b28830acee54c369/page_0007.jpg",
"images/b28830a... | In Table 4 on Page 21, what mean performance does MiCo-GPT 4 achieve? | 96.9 | float | [
16,
17,
18,
19,
20,
21,
22,
23,
24,
25,
26,
27,
28,
29
] | [
{
"page": 21,
"kind": "Table",
"description": "Table 4 lists the Mean performance for MiCo-GPT 4 as 96.9%.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/b28830acee54c369/page_0001.jpg"
},
{
"type": "image",
"image": "images/b28830acee54c369/page_0002.jpg"
},
{
"type": "image",
"image": "images/b28830acee54c369/page... |
12aaef8d1e07ad7a9a45 | 9315b6a1c80d3a75 | extract_single | en | [
"images/9315b6a1c80d3a75/page_0001.jpg",
"images/9315b6a1c80d3a75/page_0002.jpg",
"images/9315b6a1c80d3a75/page_0003.jpg",
"images/9315b6a1c80d3a75/page_0004.jpg",
"images/9315b6a1c80d3a75/page_0005.jpg",
"images/9315b6a1c80d3a75/page_0006.jpg",
"images/9315b6a1c80d3a75/page_0007.jpg",
"images/9315b6a... | According to Section 9, “Conclusie,” on Page 32, what two conclusions does the document draw about behavioural-targeting data and data protection law? | Under current law, at least according to the interpretation of the European privacy regulators, the rules for personal data generally apply to behavioural-targeting data.; Data protection law should also apply from a normative perspective. | list | [
25,
26,
27,
28,
29,
30,
31,
32
] | [
{
"page": 32,
"kind": "Text",
"description": "Section 9, “Conclusie,” states that two conclusions follow: personal-data rules generally apply to behavioural-targeting data, and data protection law should apply from a normative perspective.",
"quote": null,
"locator": null,
"support_mode": nu... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/9315b6a1c80d3a75/page_0001.jpg"
},
{
"type": "image",
"image": "images/9315b6a1c80d3a75/page_0002.jpg"
},
{
"type": "image",
"image": "images/9315b6a1c80d3a75/page... |
d9a7e43c6e3006253e21 | daeee08b06d2264b | extract_single | en | [
"images/daeee08b06d2264b/page_0001.jpg",
"images/daeee08b06d2264b/page_0002.jpg",
"images/daeee08b06d2264b/page_0003.jpg",
"images/daeee08b06d2264b/page_0004.jpg",
"images/daeee08b06d2264b/page_0005.jpg",
"images/daeee08b06d2264b/page_0006.jpg",
"images/daeee08b06d2264b/page_0007.jpg",
"images/daeee08... | According to the “IV. CONCLUSIONS” section on Page 21, how many new directions are emphasized in OED’s FY88 work program? | 4 | integer | [
13,
14,
15,
16,
17,
18,
19,
20,
21,
22,
23,
24
] | [
{
"page": 21,
"kind": "Text",
"description": "Under the section title “IV. CONCLUSIONS,” paragraph 61 states: “The four new directions emphasized in OED’s work program for FY88...”",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/daeee08b06d2264b/page_0001.jpg"
},
{
"type": "image",
"image": "images/daeee08b06d2264b/page_0002.jpg"
},
{
"type": "image",
"image": "images/daeee08b06d2264b/page... |
c67876473ffc5040a003 | 6c1b62384ade84bc | extract_single | en | [
"images/6c1b62384ade84bc/page_0001.jpg",
"images/6c1b62384ade84bc/page_0002.jpg",
"images/6c1b62384ade84bc/page_0003.jpg",
"images/6c1b62384ade84bc/page_0004.jpg",
"images/6c1b62384ade84bc/page_0005.jpg",
"images/6c1b62384ade84bc/page_0006.jpg",
"images/6c1b62384ade84bc/page_0007.jpg",
"images/6c1b623... | According to the Battery subsection of "Storage Systems" on Page 11, what main lithium battery types are listed? | Lithium Cobalt Oxide (LCO, cathode); Lithium iron phosphate (LFP, cathode); Lithium Manganese Oxide (LMO, cathode); Lithium Titanate Oxide (LTO, anode); Lithium Nickel Cobalt Aluminium oxide (NCA, cathode); Lithium Nickel Cobalt Manganese oxide (NMC, cathode) | list | [
5,
6,
7,
8,
9,
10,
11,
12,
13,
14,
15,
16,
17,
18
] | [
{
"page": 11,
"kind": "Text",
"description": "The Battery subsection under Storage Systems lists the six main lithium battery types and their electrode classifications.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/6c1b62384ade84bc/page_0001.jpg"
},
{
"type": "image",
"image": "images/6c1b62384ade84bc/page_0002.jpg"
},
{
"type": "image",
"image": "images/6c1b62384ade84bc/page... |
2b0386ab72c3675589ff | b39056eff8fec253 | extract_single | en | [
"images/b39056eff8fec253/page_0001.jpg",
"images/b39056eff8fec253/page_0002.jpg",
"images/b39056eff8fec253/page_0003.jpg",
"images/b39056eff8fec253/page_0004.jpg",
"images/b39056eff8fec253/page_0005.jpg",
"images/b39056eff8fec253/page_0006.jpg",
"images/b39056eff8fec253/page_0007.jpg",
"images/b39056e... | According to the A.4 Human Experiment Details section on Page 17, how many cold-start rounds does each task use? | 4 | integer | [
11,
12,
13,
14,
15,
16,
17,
18,
19,
20,
21
] | [
{
"page": 17,
"kind": "Text",
"description": "In the A.4 Human Experiment Details section, the cold-start stage states that each task uses four cold-start rounds.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/b39056eff8fec253/page_0001.jpg"
},
{
"type": "image",
"image": "images/b39056eff8fec253/page_0002.jpg"
},
{
"type": "image",
"image": "images/b39056eff8fec253/page... |
db9ddf505149811cbad9 | 1a5ad50836173e08 | extract_single | en | [
"images/1a5ad50836173e08/page_0001.jpg",
"images/1a5ad50836173e08/page_0002.jpg",
"images/1a5ad50836173e08/page_0003.jpg",
"images/1a5ad50836173e08/page_0004.jpg",
"images/1a5ad50836173e08/page_0005.jpg",
"images/1a5ad50836173e08/page_0006.jpg",
"images/1a5ad50836173e08/page_0007.jpg",
"images/1a5ad50... | According to Section 3.2, Interviews, on Page 7, how many participants were interviewed? | 11 | integer | [
5,
6,
7,
8,
9,
10,
11,
12,
13,
14,
15,
16,
17
] | [
{
"page": 7,
"kind": "Text",
"description": "The section 3.2 Interviews states that the researchers conducted interviews with 11 participants.",
"quote": null,
"locator": null,
"support_mode": null
}
] | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/1a5ad50836173e08/page_0001.jpg"
},
{
"type": "image",
"image": "images/1a5ad50836173e08/page_0002.jpg"
},
{
"type": "image",
"image": "images/1a5ad50836173e08/page... |
4521a46e257c95eda548 | 74f2f8a1351dc3d5 | extract_multi | en | [
"images/74f2f8a1351dc3d5/page_0001.jpg",
"images/74f2f8a1351dc3d5/page_0002.jpg",
"images/74f2f8a1351dc3d5/page_0003.jpg",
"images/74f2f8a1351dc3d5/page_0004.jpg",
"images/74f2f8a1351dc3d5/page_0005.jpg",
"images/74f2f8a1351dc3d5/page_0006.jpg",
"images/74f2f8a1351dc3d5/page_0007.jpg",
"images/74f2f8a... | In the section “Generalized gradients as limiting Fréchet subgradients,” what does Proposition 3.12 establish, and what equality does Corollary 3.13 give under its C¹ condition on Pages 32 and 33? | Proposition 3.12 establishes that G(θ) belongs to the limiting Fréchet subgradient set of L₀ at θ.; Corollary 3.13 establishes that G(θ) = (∇L₀)(θ) for θ in an open set U where L₀|ᵤ is C¹(U, ℝ). | list | [
30,
31,
32,
33,
34,
35,
36,
37,
38
] | [
{
"page": 32,
"kind": "Text",
"description": "Under Proposition 3.12 in Section 3.7, the text concludes that G(θ) ∈ (𝔻L₀)(θ), the limiting Fréchet subgradient set.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 33,
"kind": "Text",
"description": "Under C... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/74f2f8a1351dc3d5/page_0001.jpg"
},
{
"type": "image",
"image": "images/74f2f8a1351dc3d5/page_0002.jpg"
},
{
"type": "image",
"image": "images/74f2f8a1351dc3d5/page... |
14ecd40e034d26069228 | 8b48ef912ceceecd | extract_multi | en | [
"images/8b48ef912ceceecd/page_0001.jpg",
"images/8b48ef912ceceecd/page_0002.jpg",
"images/8b48ef912ceceecd/page_0003.jpg",
"images/8b48ef912ceceecd/page_0004.jpg",
"images/8b48ef912ceceecd/page_0005.jpg",
"images/8b48ef912ceceecd/page_0006.jpg",
"images/8b48ef912ceceecd/page_0007.jpg",
"images/8b48ef9... | According to the section “Analyzing pathways to reasoning strategies: Data selection and interpolation” on Pages 8 and 9, what two conclusions are drawn about the effects of training-data format and model merging on reasoning strategies? | Training data format has a greater impact on reasoning strategies than domain.; Model merging interpolates reasoning strategies from multiple-choice-trained models toward free-form-trained models without additional training. | list | [
3,
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14,
15,
16,
17
] | [
{
"page": 8,
"kind": "Text",
"description": "In Section 5.1, format variation is reported to produce larger shifts in reasoning strategies than domain differences.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 9,
"kind": "Text",
"description": "In Sectio... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/8b48ef912ceceecd/page_0001.jpg"
},
{
"type": "image",
"image": "images/8b48ef912ceceecd/page_0002.jpg"
},
{
"type": "image",
"image": "images/8b48ef912ceceecd/page... |
6f9698de3a893e687d03 | 56d7426c35a6cb77 | extract_multi | en | [
"images/56d7426c35a6cb77/page_0001.jpg",
"images/56d7426c35a6cb77/page_0002.jpg",
"images/56d7426c35a6cb77/page_0003.jpg",
"images/56d7426c35a6cb77/page_0004.jpg",
"images/56d7426c35a6cb77/page_0005.jpg",
"images/56d7426c35a6cb77/page_0006.jpg",
"images/56d7426c35a6cb77/page_0007.jpg",
"images/56d7426... | According to the HIV-1 Protease discussion on Pages 7 and 21, what proportions are reported for the closed/semi-open states in (a) the base Boltz-2 samples and (b) the two-component Gaussian-mixture target? | 83% closed and 17% semi-open; 25% closed and 75% semi-open | list | [
7,
8,
9,
10,
11,
12,
13,
14,
15,
16,
17,
18,
19,
20,
21
] | [
{
"page": 7,
"kind": "Text",
"description": "In the HIV-1 Protease discussion, the base Boltz-2 model samples states with proportions of 83% closed and 17% semi-open.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 21,
"kind": "Text",
"description": "The H... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/56d7426c35a6cb77/page_0001.jpg"
},
{
"type": "image",
"image": "images/56d7426c35a6cb77/page_0002.jpg"
},
{
"type": "image",
"image": "images/56d7426c35a6cb77/page... |
207ba394afa40bc74fa4 | bcf8d2a961986f46 | extract_multi | en | [
"images/bcf8d2a961986f46/page_0001.jpg",
"images/bcf8d2a961986f46/page_0002.jpg",
"images/bcf8d2a961986f46/page_0003.jpg",
"images/bcf8d2a961986f46/page_0004.jpg",
"images/bcf8d2a961986f46/page_0005.jpg",
"images/bcf8d2a961986f46/page_0006.jpg",
"images/bcf8d2a961986f46/page_0007.jpg",
"images/bcf8d2a... | According to Appendix C on Page 28 and the application in Appendix D on Pages 30–31, what variance order is used for I₁ and what exponential upper bound does Bernstein’s inequality yield for P(∑_{i=1}^n I_{1,i} ≥ μnI(X₁;Y₂|X₂,U))? | The variance of I₁ is O(n^{-1/2}).; P(∑_{i=1}^n I_{1,i} ≥ μnI(X₁;Y₂|X₂,U)) ≤ exp(−ζ₁n^{1/2}) for some ζ₁ > 0. | list | [
24,
25,
26,
27,
28,
29,
30,
31,
32,
33,
34,
35
] | [
{
"page": 28,
"kind": "Text",
"description": "Appendix C states Bernstein’s inequality for independent zero-mean bounded random variables, which is recalled for the application.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 30,
"kind": "Text",
"descripti... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/bcf8d2a961986f46/page_0001.jpg"
},
{
"type": "image",
"image": "images/bcf8d2a961986f46/page_0002.jpg"
},
{
"type": "image",
"image": "images/bcf8d2a961986f46/page... |
85b99cd80e80dc0d0ed7 | bfcda28cc47ecf68 | extract_multi | en | [
"images/bfcda28cc47ecf68/page_0001.jpg",
"images/bfcda28cc47ecf68/page_0002.jpg",
"images/bfcda28cc47ecf68/page_0003.jpg",
"images/bfcda28cc47ecf68/page_0004.jpg",
"images/bfcda28cc47ecf68/page_0005.jpg",
"images/bfcda28cc47ecf68/page_0006.jpg",
"images/bfcda28cc47ecf68/page_0007.jpg",
"images/bfcda28... | In the Experiments sections “Temporal Generalisation” and “Phenological alignment,” what are the mean mIoU values for U-TAE and TSViT in Table 3, and what are their corresponding mIoU values using T³S in Table 4? | U-TAE: 35.8% (Table 3 mean mIoU); 34.6% (Table 4, T³S); TSViT: 48.1% (Table 3 mean mIoU); 37.4% (Table 4, T³S) | list | [
5,
6,
7,
8,
9,
10,
11,
12,
13,
14
] | [
{
"page": 11,
"kind": "Table",
"description": "Table 3 in the Temporal Generalisation results reports mean mIoU values of 35.8% for U-TAE and 48.1% for TSViT.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 12,
"kind": "Table",
"description": "Table 4 in t... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/bfcda28cc47ecf68/page_0001.jpg"
},
{
"type": "image",
"image": "images/bfcda28cc47ecf68/page_0002.jpg"
},
{
"type": "image",
"image": "images/bfcda28cc47ecf68/page... |
edea7998c172ebd13a51 | 483f2c8d24ff4aa5 | extract_multi | en | [
"images/483f2c8d24ff4aa5/page_0001.jpg",
"images/483f2c8d24ff4aa5/page_0002.jpg",
"images/483f2c8d24ff4aa5/page_0003.jpg",
"images/483f2c8d24ff4aa5/page_0004.jpg",
"images/483f2c8d24ff4aa5/page_0005.jpg",
"images/483f2c8d24ff4aa5/page_0006.jpg",
"images/483f2c8d24ff4aa5/page_0007.jpg",
"images/483f2c8... | According to Section 7.2 on Pages 30–31, on how many markets did an adaptive gain attain the lowest mean negative log score, and on how many markets was the pooled DMD-logit-minus-constant differential negative? | 12; 11 | list | [
30,
31,
32,
33,
34,
35,
36,
37,
38,
39,
40
] | [
{
"page": 30,
"kind": "Text",
"description": "In Section 7.2, the text states that an adaptive gain attains the lowest mean negative log score on every one of the twelve markets.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 31,
"kind": "Text",
"descript... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/483f2c8d24ff4aa5/page_0001.jpg"
},
{
"type": "image",
"image": "images/483f2c8d24ff4aa5/page_0002.jpg"
},
{
"type": "image",
"image": "images/483f2c8d24ff4aa5/page... |
3f440746d99f88dd545a | f07e40ccc25fba45 | extract_multi | en | [
"images/f07e40ccc25fba45/page_0001.jpg",
"images/f07e40ccc25fba45/page_0002.jpg",
"images/f07e40ccc25fba45/page_0003.jpg",
"images/f07e40ccc25fba45/page_0004.jpg",
"images/f07e40ccc25fba45/page_0005.jpg",
"images/f07e40ccc25fba45/page_0006.jpg",
"images/f07e40ccc25fba45/page_0007.jpg",
"images/f07e40c... | According to the PRISON framework on Page 3 and the Experiment 1 Results on Page 7, what are the five criminal trait dimensions, and which trait emerged as the dominant strategy? | False Statements; Frame-Up; Psychological Manipulation; Emotional Disguise; Moral Disengagement; Psychological Manipulation was the dominant strategy | list | [
3,
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14,
15
] | [
{
"page": 3,
"kind": "Text",
"description": "In the PRISON section’s Criminal Traits subsection, the five core trait dimensions are listed as False Statements, Frame-Up, Psychological Manipulation, Emotional Disguise, and Moral Disengagement.",
"quote": null,
"locator": null,
"support_mode":... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/f07e40ccc25fba45/page_0001.jpg"
},
{
"type": "image",
"image": "images/f07e40ccc25fba45/page_0002.jpg"
},
{
"type": "image",
"image": "images/f07e40ccc25fba45/page... |
db5330dd16fef8c44fa7 | 5bb686fdf504488e | extract_multi | en | [
"images/5bb686fdf504488e/page_0001.jpg",
"images/5bb686fdf504488e/page_0002.jpg",
"images/5bb686fdf504488e/page_0003.jpg",
"images/5bb686fdf504488e/page_0004.jpg",
"images/5bb686fdf504488e/page_0005.jpg",
"images/5bb686fdf504488e/page_0006.jpg",
"images/5bb686fdf504488e/page_0007.jpg",
"images/5bb686f... | According to the RQ2 section on Pages 17–18, what were the four reported maximum metrics for heterogeneous composition? | Maximum trajectory RMSE: 1.28 × 10^-16; Maximum mass-conservation error: 6.02 × 10^-16; Maximum event-order mismatch: 0; Maximum port-type violation: 0 | list | [
10,
11,
12,
13,
14,
15,
16,
17,
18
] | [
{
"page": 17,
"kind": "Text",
"description": "In the RQ2 discussion, the trajectory RMSE and conservation error are reported at floating-point tolerance, with zero event-order and port-type violations.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 18,
"kind"... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/5bb686fdf504488e/page_0001.jpg"
},
{
"type": "image",
"image": "images/5bb686fdf504488e/page_0002.jpg"
},
{
"type": "image",
"image": "images/5bb686fdf504488e/page... |
431e253bff292a9c090d | b38270c3a17fb6ef | extract_multi | en | [
"images/b38270c3a17fb6ef/page_0001.jpg",
"images/b38270c3a17fb6ef/page_0002.jpg",
"images/b38270c3a17fb6ef/page_0003.jpg",
"images/b38270c3a17fb6ef/page_0004.jpg",
"images/b38270c3a17fb6ef/page_0005.jpg",
"images/b38270c3a17fb6ef/page_0006.jpg",
"images/b38270c3a17fb6ef/page_0007.jpg",
"images/b38270c... | Based on Graph 2 on Page 18 and the “Human capacity—Development in early childhood” discussion on Page 24, what two conclusions connect early cognitive differences with equality of opportunity? | Cognitive development scores differ notably by family background between ages three and five.; Investments in early childhood can be decisive for improving equality of opportunity. | list | [
13,
14,
15,
16,
17,
18,
19,
20,
21,
22,
23,
24
] | [
{
"page": 18,
"kind": "Figure",
"description": "Graph 2 states that the cognitive development of children aged three to five differs notably according to family background.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 24,
"kind": "Text",
"description": ... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/b38270c3a17fb6ef/page_0001.jpg"
},
{
"type": "image",
"image": "images/b38270c3a17fb6ef/page_0002.jpg"
},
{
"type": "image",
"image": "images/b38270c3a17fb6ef/page... |
bb93feb7e39ca2676d3f | 588c85a0e9689f49 | extract_multi | en | [
"images/588c85a0e9689f49/page_0001.jpg",
"images/588c85a0e9689f49/page_0002.jpg",
"images/588c85a0e9689f49/page_0003.jpg",
"images/588c85a0e9689f49/page_0004.jpg",
"images/588c85a0e9689f49/page_0005.jpg",
"images/588c85a0e9689f49/page_0006.jpg",
"images/588c85a0e9689f49/page_0007.jpg",
"images/588c85a... | According to Proposition 6.1 on page 38 and the discussion on page 41, which theories are established as supported malleable through skew commutativity and skew associativity, and which theory requires a separate proof of malleability? | SL, GA, and CA are supported malleable via skew commutativity and skew associativity.; GC requires a separate proof of malleability. | list | [
33,
34,
35,
36,
37,
38,
39,
40,
41
] | [
{
"page": 38,
"kind": "Text",
"description": "Proposition 6.1 states that an equational theory is malleable when it satisfies skew commutativity and skew associativity.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 41,
"kind": "Text",
"description": "The... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/588c85a0e9689f49/page_0001.jpg"
},
{
"type": "image",
"image": "images/588c85a0e9689f49/page_0002.jpg"
},
{
"type": "image",
"image": "images/588c85a0e9689f49/page... |
6fdd9ce9d396d533b263 | ea8791b72cbc11d8 | extract_multi | en | [
"images/ea8791b72cbc11d8/page_0001.jpg",
"images/ea8791b72cbc11d8/page_0002.jpg",
"images/ea8791b72cbc11d8/page_0003.jpg",
"images/ea8791b72cbc11d8/page_0004.jpg",
"images/ea8791b72cbc11d8/page_0005.jpg",
"images/ea8791b72cbc11d8/page_0006.jpg",
"images/ea8791b72cbc11d8/page_0007.jpg",
"images/ea8791b... | Based on the “Summary of contributions” on Page 2 and Theorem VI.7 in the quantum relational Hoare logic section on Page 6, for what class of programs and postconditions is completeness claimed? | AST (almost-surely terminating) programs with bounded postconditions | string | [
2,
3,
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14,
15,
16
] | [
{
"page": 2,
"kind": "Text",
"description": "The “Summary of contributions” states that the relational program logic is sound and complete for quantum programs, citing Theorems VI.3 and VI.7.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 6,
"kind": "Text",
... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/ea8791b72cbc11d8/page_0001.jpg"
},
{
"type": "image",
"image": "images/ea8791b72cbc11d8/page_0002.jpg"
},
{
"type": "image",
"image": "images/ea8791b72cbc11d8/page... |
747832d8fd7ab846286e | 9cea1327ea50ce28 | extract_multi | en | [
"images/9cea1327ea50ce28/page_0001.jpg",
"images/9cea1327ea50ce28/page_0002.jpg",
"images/9cea1327ea50ce28/page_0003.jpg",
"images/9cea1327ea50ce28/page_0004.jpg",
"images/9cea1327ea50ce28/page_0005.jpg",
"images/9cea1327ea50ce28/page_0006.jpg",
"images/9cea1327ea50ce28/page_0007.jpg",
"images/9cea132... | Based on Theorem 1.2 on Page 4 and Theorem 3.2 on Page 14, what are the stated running-time guarantee and inapproximability result? | An optimal orientation can be computed in time n^{O(k)}; Unless P = NP, no polynomial-time algorithm can achieve a 581773/581774-approximation | list | [
1,
2,
3,
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14
] | [
{
"page": 4,
"kind": "Text",
"description": "Theorem 1.2 states that an optimal orientation of an MMRO instance can be computed in time n^{O(k)}.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 14,
"kind": "Text",
"description": "Theorem 3.2 states that, u... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/9cea1327ea50ce28/page_0001.jpg"
},
{
"type": "image",
"image": "images/9cea1327ea50ce28/page_0002.jpg"
},
{
"type": "image",
"image": "images/9cea1327ea50ce28/page... |
7ca0632de96bfc74b504 | 337779e1266d4b73 | extract_multi | en | [
"images/337779e1266d4b73/page_0001.jpg",
"images/337779e1266d4b73/page_0002.jpg",
"images/337779e1266d4b73/page_0003.jpg",
"images/337779e1266d4b73/page_0004.jpg",
"images/337779e1266d4b73/page_0005.jpg",
"images/337779e1266d4b73/page_0006.jpg",
"images/337779e1266d4b73/page_0007.jpg",
"images/337779e... | According to the “Spotlight on Cambodia” pages (Pages 25 and 26), which contracting approach increased health-service use among the poorest half by nearly 30 percentage points, and what verification safeguard did its contracts require? | Contracting out; An independent performance verification system | list | [
20,
21,
22,
23,
24,
25,
26,
27,
28,
29,
30
] | [
{
"page": 25,
"kind": "Text",
"description": "In the Cambodia spotlight, use among the poorest half of the population increased by nearly 30 percentage points in contracted-out districts.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 26,
"kind": "Text",
... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/337779e1266d4b73/page_0001.jpg"
},
{
"type": "image",
"image": "images/337779e1266d4b73/page_0002.jpg"
},
{
"type": "image",
"image": "images/337779e1266d4b73/page... |
bf4fdf548972e30353b0 | 219fe939092ec6d4 | extract_multi | en | [
"images/219fe939092ec6d4/page_0001.jpg",
"images/219fe939092ec6d4/page_0002.jpg",
"images/219fe939092ec6d4/page_0003.jpg",
"images/219fe939092ec6d4/page_0004.jpg",
"images/219fe939092ec6d4/page_0005.jpg",
"images/219fe939092ec6d4/page_0006.jpg",
"images/219fe939092ec6d4/page_0007.jpg",
"images/219fe93... | According to the “Analysis of Thought Styles” section on Pages 16–17, what were the 1-shot hidden-thoughts accuracies and average MMLU inference thinking-token counts for the STEM and Law models? | STEM: 76.3% accuracy; 125.7 average inference thinking tokens; Law: 72.0% accuracy; 143.3 average inference thinking tokens | list | [
9,
10,
11,
12,
13,
14,
15,
16,
17
] | [
{
"page": 16,
"kind": "Table",
"description": "Table 3 reports 1-shot hidden-thoughts accuracies of 76.3% for Reasoning CPT (STEM) and 72.0% for Reasoning CPT (Law).",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 17,
"kind": "Table",
"description": "Table... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/219fe939092ec6d4/page_0001.jpg"
},
{
"type": "image",
"image": "images/219fe939092ec6d4/page_0002.jpg"
},
{
"type": "image",
"image": "images/219fe939092ec6d4/page... |
c48f976d02649db419bd | ca044633034b2899 | extract_multi | en | [
"images/ca044633034b2899/page_0001.jpg",
"images/ca044633034b2899/page_0002.jpg",
"images/ca044633034b2899/page_0003.jpg",
"images/ca044633034b2899/page_0004.jpg",
"images/ca044633034b2899/page_0005.jpg",
"images/ca044633034b2899/page_0006.jpg",
"images/ca044633034b2899/page_0007.jpg",
"images/ca04463... | In the “Absolute binding free energy” section, based on Pages 44 and 48, what four Boltz2 affinity values were used for compound selection, and what Pearson correlation was reported between Boltz2 and AFEP estimated affinities? | 5, 6, 7, and 8 pKi, each with a ±0.25 range; 0.53 | list | [
40,
41,
42,
43,
44,
45,
46,
47,
48,
49
] | [
{
"page": 44,
"kind": "Text",
"description": "Under “Absolute binding free energy,” compounds were selected around four Boltz2 estimated affinity values: 5, 6, 7, and 8, each within ±0.25.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 48,
"kind": "Figure",
... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/ca044633034b2899/page_0001.jpg"
},
{
"type": "image",
"image": "images/ca044633034b2899/page_0002.jpg"
},
{
"type": "image",
"image": "images/ca044633034b2899/page... |
a53f175bb6118d90d91d | b58c0624943cb5ce | extract_multi | en | [
"images/b58c0624943cb5ce/page_0001.jpg",
"images/b58c0624943cb5ce/page_0002.jpg",
"images/b58c0624943cb5ce/page_0003.jpg",
"images/b58c0624943cb5ce/page_0004.jpg",
"images/b58c0624943cb5ce/page_0005.jpg",
"images/b58c0624943cb5ce/page_0006.jpg",
"images/b58c0624943cb5ce/page_0007.jpg",
"images/b58c062... | In the “Gradient estimation” section, based on Pages 15 and 16, how does gradient-estimation time scale for weight versus node perturbation, and what effect does task complexity have? | Weight perturbation: proportional to the total number of network parameters (N); Node perturbation: proportional to the square root of the total number of network parameters (√N); Task complexity had no effect on gradient estimation | list | [
10,
11,
12,
13,
14,
15,
16,
17
] | [
{
"page": 15,
"kind": "Figure",
"description": "Figure 3 caption states that the expected scaling is N for weight perturbation and √N for node perturbation, and panel (c) shows gradient-estimation time across task complexities.",
"quote": null,
"locator": null,
"support_mode": null
},
{
... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/b58c0624943cb5ce/page_0001.jpg"
},
{
"type": "image",
"image": "images/b58c0624943cb5ce/page_0002.jpg"
},
{
"type": "image",
"image": "images/b58c0624943cb5ce/page... |
47825d7d0e60ebcb4d08 | ef96d6a7188cfd46 | extract_multi | en | [
"images/ef96d6a7188cfd46/page_0001.jpg",
"images/ef96d6a7188cfd46/page_0002.jpg",
"images/ef96d6a7188cfd46/page_0003.jpg",
"images/ef96d6a7188cfd46/page_0004.jpg",
"images/ef96d6a7188cfd46/page_0005.jpg",
"images/ef96d6a7188cfd46/page_0006.jpg",
"images/ef96d6a7188cfd46/page_0007.jpg",
"images/ef96d6a... | Based on Table 4 on Page 21 and the “Self-supervised and intrinsic quality objectives” discussion on Page 22, what are the fine-grained dimension counts for the Wachsmuth et al. and Ivanova et al. taxonomies, and what correlation does Intrinsic QA achieve against crowdsourced judgments? | Wachsmuth et al.: 15; Ivanova et al.: 15; intrinsic QA correlation: 0.46 | string | [
15,
16,
17,
18,
19,
20,
21,
22,
23,
24
] | [
{
"page": 21,
"kind": "Table",
"description": "Table 4 lists 15 fine-grained dimensions for both the Wachsmuth et al. (2017a) and Ivanova et al. (2024) taxonomies.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 22,
"kind": "Text",
"description": "The disc... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/ef96d6a7188cfd46/page_0001.jpg"
},
{
"type": "image",
"image": "images/ef96d6a7188cfd46/page_0002.jpg"
},
{
"type": "image",
"image": "images/ef96d6a7188cfd46/page... |
cc2be6fb9279bba4d03e | 2d3d973901dcc581 | extract_multi | en | [
"images/2d3d973901dcc581/page_0001.jpg",
"images/2d3d973901dcc581/page_0002.jpg",
"images/2d3d973901dcc581/page_0003.jpg",
"images/2d3d973901dcc581/page_0004.jpg",
"images/2d3d973901dcc581/page_0005.jpg",
"images/2d3d973901dcc581/page_0006.jpg",
"images/2d3d973901dcc581/page_0007.jpg",
"images/2d3d973... | According to Table 5 on Page 21 and Table 6 on Page 25, what capability coverage is assigned to the Code and science teacher, and which benchmarks are listed under Long Context and Instruction Following? | Program synthesis, scientific computing, and physical reasoning; AA-LCR; IFBench | list | [
20,
21,
22,
23,
24,
25,
26,
27,
28,
29,
30
] | [
{
"page": 21,
"kind": "Table",
"description": "Table 5 identifies the Code and science teacher's coverage as program synthesis, scientific computing, and physical reasoning.",
"quote": null,
"locator": null,
"support_mode": null
},
{
"page": 25,
"kind": "Table",
"description"... | {
"verdict": "pass",
"scores": {
"groundedness": 4,
"answer_correctness": 4,
"question_clarity": 4,
"evidence_sufficiency": 4,
"task_fidelity": 4,
"difficulty": 3
},
"issue_codes": [],
"verifier_model": "gpt-5.6-luna",
"verifier_prompt_version": "mmprolong-legacy-verifier-v2",
"rep... | [
{
"role": "user",
"content": [
{
"type": "image",
"image": "images/2d3d973901dcc581/page_0001.jpg"
},
{
"type": "image",
"image": "images/2d3d973901dcc581/page_0002.jpg"
},
{
"type": "image",
"image": "images/2d3d973901dcc581/page... |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
MMProLong verified preview
This review-only subset contains 30 samples selected from server_3000_v2.
All records passed the v2 hard gates and independent verifier. Each document is copied once,
and train_qwen.jsonl references the complete rendered page sequence with repository-relative paths.
This is a preview artifact, not the final public release.
- Downloads last month
- 8