Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

Where V2 differs from V1

Every annotation in data/samples.json that differs from upstream MMLongBench-Doc, and why. 106 of v1's 1,082 questions are changed; the rest are byte-identical to upstream.

Most were found the same way: a document-QA system answered the question, scored zero, and the answer turned out on inspection to be right or at least as defensible as the key. Each was then checked against the source PDF, and the check is written into the entry's note — the page, the figures, and the arithmetic — so a reader can disagree with the reasoning rather than just the verdict.

One structural change: v1 contains the same question twice under PRE_2022.09.29_NSL-politics_REPORT.pdf, once answered 46% and once Not answerable. Page 97 answers it, so both were resolved to 46%, which left a duplicate (doc_id, question) key — enough to silently drop a row from any keyed join. The second copy is removed. Ten further rows are dropped as corpus_defect (below), so V2 has 1,071 questions over 134 documents, to v1's 1,082 over 135.

Files

File What it is
corrections.json The 106 changes, each with the original annotation, the correction, and a note giving the evidence.
diff.json The same changes as a plain field-level before/after, no commentary.
samples_v1.json Upstream's original samples.json, unmodified, so the diff can be reproduced independently.

Reproduce the diff yourself — join on (doc_id, question) rather than by position, since V2 drops 11 rows (one duplicate, ten corpus_defect) and rewords 38 questions:

import json
norm = lambda s: " ".join(str(s).split())
v1 = {(x["doc_id"], norm(x["question"])): x for x in json.load(open("data/corrections/samples_v1.json"))}
v2 = {(x["doc_id"], norm(x["question"])): x for x in json.load(open("data/samples.json"))}
corr = json.load(open("data/corrections/corrections.json"))

for c in corr:                              # each correction names both sides of the rename
    before = v1[(c["doc_id"], norm(c["question"]))]
    after  = v2.get((c["doc_id"], norm(c.get("manual_question") or c["question"])))
    if after is None:                       # corpus_defect: dropped from V2 entirely
        print(c["doc_id"], c["dispute_type"], "removed")
        continue
    changed = {k for k in ("answer", "answer_format", "evidence_pages")
               if str(before[k]) != str(after[k])}
    print(c["doc_id"], c["dispute_type"], changed or "question only")

By category

dispute_type n Meaning
wrong_answer 26 The question is sound, but the recorded answer is not what the document says.
defective_question 19 The question itself is broken — wrong page, wrong year, a typo, a false premise, or singular phrasing over a multi-part answer. V2 rewords it and keeps the answer.
ambiguous_question 19 Two readings are both defensible and the recorded answer only fits one. V2 states the intended reading in the question.
should_be_answerable 14 Labelled Not answerable, but the document contains a determinate answer.
incomplete_answer 16 The recorded key omits members, or omits a form of the answer that is equally correct.
corpus_defect 10 No answer is recoverable from the distributed corpus at all. Removed from samples.json rather than counted wrong.
should_be_unanswerable 2 An answer is recorded, but the document does not support one.

By field

Field Rows changed
answer 74
question 38
evidence_pages 35
answer_format 19

corpus_defect rows are not present in samples.json — they are listed below and in corrections.json as the record of why they were dropped. Their manual_answer is None because no answer is recoverable, not because the answer is empty. All ten belong to one document, whose distributed PDF is a different deck from the one the questions were written against, so V2 covers 134 of upstream's 135 PDFs. eval/metrics.py still filters on the flag, which makes predictions produced against the earlier 1,081-row file score correctly rather than counting those rows as failures.

Fourteen Not answerable keys were widened to Not answerable or 0 or none. The document does let you determine that the set is empty in each case, so reporting "there are none" — or, for a how many, the count 0 — says exactly what the key means. eval/metrics.py matches the Not answerable prefix, so these rows still count in the unanswerable slice.

This was needed because the corpus keys the same shape two ways. Of the 287 upstream questions beginning "how many", 14 are keyed 0 for a thing that is simply absent — cats, tigers, airplanes, blue arrows, GPT-4o, words starting with 'X' — while 56 are keyed Not answerable, including questions of essentially identical shape. One brochure contradicts itself: GPL-Graduate-Studies keys "how many dogs and cats are there in page 17" as ['0','0'] and "how many people with scarf are there in Page 5" as 2, but the sunglasses twin on the same page as Not answerable. It cuts both ways — RAR.pdf "how many GPT-4o examples" is keyed 0 (the string appears on none of its 28 pages; the paper uses GPT-4V), so a system that declines there is marked wrong too. Widening the key settles it without deleting a negative sample in either direction.

The widening is not applied to the other how many rows keyed Not answerable. Those fail for different reasons: page ranges outside the document (Pages 200-205, Pages 400-640), entities the document never covers (Boeing, Airbus, Shopee, HAQ Executive Leadership, PWC Technic), or deliberate terminology twins (EFI for ECU, NTUSU for NUSSU, generation for correction).

When a Not answerable key gets widened

All 208 Not answerable rows were reviewed against this test, which turns on what is missing:

Widen when the container the question names exists and the thing it asks about is verifiably absent from it. Then "none" — or, for a how many, 0 — is a reading of the document, not a guess, and it means exactly what the key means. Examples: page 2 of 8e7c4cb… has a street photograph with three cars and a box truck and no bicycle; PDF p10 and p14 of e79deb02… contain zero images and zero large vector blocks, so there is no diagram on "page 10" under either page numbering; the timer chart on p50 of User_Manual_1500S draws only green and blue bars, so no red bar starts anywhere.

Do not widen in four cases, where "none" is an over-claim rather than a reading:

Case Example
The container itself is missing Pages 200-205, Pages 400-640; 2311.16502v3.pdf "Figure 10", which has no caption anywhere in the paper
The entity never appears in the document Beijing in a Nepal media survey; a microwave in a dishwasher manual; TRUCS "level-6"; survey years 2018 and 2022
The question asks for a complement against an unbounded universe mi_phone.pdf, "signal icons that can not be found in Status Bar"
The set is short, not empty the EarthLink deck asks for "another two companies" when exactly one further contact carries a phone number

The last row is why this is a test and not a heuristic: it looked like an empty set until the contact slide was actually read.

Every change

marks a question whose wording was changed; the answer column then shows the answer that wording now has.

Document Question (v1) V1 answer V2 answer Type
2005.12872v3.pdf How many multi-head self-attention layers does DETR's d… 36 12 wrong_answer
2005.12872v3.pdf According to Fig 10 architecture of DETR’s transformer,… Red Pink wrong_answer
2023.findings-emnlp.248.pdf Which dataset used in this paper was proposed in 2022 a… ProofWriter AR-LSAT wrong_answer
698bba535087fa9a7f9009e172a7f763.… What are the counties mentioned in the document? ['Hamilton', 'Lucas', '… ['Hamilton', 'Lucas', '… wrong_answer
91521110100M_4K_UHD_Display_User_… A transmitter operating at a frequency of 500 MHz has a… 2.92 3.80 wrong_answer
ADOBE_2015_10K.pdf what is roa for ADBE in FY2015? 0.053 0.054 wrong_answer
AMAZON_2017_10K.pdf what is the percentage change of return for allowance f… 60.3% -60.3% wrong_answer
AMAZON_2017_10K.pdf What is Amazon's FY2017 days payable outstanding (DPO)?… 97.75 97.70 wrong_answer
BESTBUY_2023_10K.pdf what is the change of Best Buy's gross margins change f… 1.08% 0.12% wrong_answer
GPL-Graduate-Studies-Professional… Which programme by coursework with disciplinary content… ['MA (Humanities Educat… ['MA (Humanities Educat… wrong_answer
NUS-FASS-Graduate-Guidebook-2021-… Which of the following department does not provide a Ph… Department of History None wrong_answer
PIP_Seniors-and-Tech-Use_040314.p… What is the gap between male 65+ age group who use inte… 73.0 12.0 wrong_answer
PIP_Seniors-and-Tech-Use_040314.p… How many people who do not go online or only use SNS in… 4087 1038 wrong_answer
PIP_Seniors-and-Tech-Use_040314.p… How many 65+ age group people go online 3-5 times per w… 1251 738 wrong_answer
PI_2017.10.04_Automation_FINAL.pdf How many US workers say email or social media have had … 2481 1506 wrong_answer
PI_2018.11.19_algorithms_FINAL.pdf Among all interviewees in the survey, what percentage o… 20% 5.1% wrong_answer
PRE_2022.09.29_NSL-politics_REPOR… Which Hispanic origin group in the United States is mos… Puerto Rican Cuban wrong_answer
afe620b9beac86c1027b96d31d396407.… What are the bankers' names associated with GODFREY PHI… ['State Bank of India',… ['State Bank of India',… wrong_answer
afe620b9beac86c1027b96d31d396407.… For the year ended March 31,2003, how much less were th… 83672770 8367277000 wrong_answer
amb-siteaudits-ds15-150204174043-… What is the Top-Level Page name of the page with the sl… /video/videocat/video92… /category3/subcat2/ wrong_answer
bdf54dxa.pdf How many possible problems does the diswasher may encou… 17 15 wrong_answer
e79deb02a0c0e87511080836c5d4347b.… How many strengths and weaknesses are metioned in Appen… ['23', '21'] ['22', '21'] wrong_answer
f8d3a162ab9507e021d83dd109118b60.… How many quizzes are there in the entire course? 6 7 wrong_answer
germanwingsdigitalcrisisanalysis-… Is the Germanwings Facebook account logo consistent bef… yes no wrong_answer
germanwingsdigitalcrisisanalysis-… When did the number of tweets referencing Germanwings e… 14:04 CET 13:51 CET wrong_answer
obs-productdesc-en.pdf What is the benefit of level-2 in the system has passed… Multi-AZ storage Erasure Code wrong_answer
05-03-18-political-release.pdf What is the percentage of registered voters who support… ⟳ 92% 92% defective_question
05-03-18-political-release.pdf How many Demoncratic people in the survey of U.S. adult… ⟳ 128 128 defective_question
05-03-18-political-release.pdf How many non-partisan people in the survey of U.S. adul… ⟳ Not answerable Not answerable defective_question
11-21-16-Updated-Post-Election-Re… How many % of voters reactions are "uneasy" and "excite… ⟳ [53, 1.4] [53, 51] defective_question
2021-Apple-Catalog.pdf One40 can only be used for Apple Watch, is that true? P… ⟳ Yes No defective_question
2311.16502v3.pdf According to this paper, among nice different datasets … ⟳ "MMMU" "MMMU" defective_question
2401.18059v1.pdf Write down the pseudo code from appendix that correspon… ⟳ Slayer ← sorted(top k)[… Slayer ← sorted(top k)[… defective_question
AMAZON_2017_10K.pdf How do Amazon recognize least cost? ⟳ straight-line basis wit… straight-line basis wit… defective_question
PG_2020.05.21_International-Coope… How many EU people believe that they will have more inf… ⟳ 19% 19% defective_question
PI_2017.10.04_Automation_FINAL.pdf How many US workers are interested in a robot caregiver… ⟳ 1695 1695 defective_question
SnapNTell.pdf In the 3rd Wiki filtering, how many more entities were … ⟳ 2885 1923 defective_question
User_Manual_1500S_Classic_EN.pdf How many steps are there for data exchange via USB? ⟳ 9 9 defective_question
afe620b9beac86c1027b96d31d396407.… How much higher was the proposed dividend paid (Rupees … ⟳ 155.98 155.98 defective_question
b3m5kaeqm2w8n4bwcesw-140602121350… Does CFCs causes skin burn? Directly answer 'yes' or 'n… ⟳ Yes No defective_question
e79deb02a0c0e87511080836c5d4347b.… What are the words written in the first rectangle on th… ⟳ ['strategic priority ar… ['strategic priority ar… defective_question
earthlinkweb-150213112111-convers… What is the job of the contact person in the picture at… ⟳ Vice President of Produ… Vice President of Produ… defective_question
f1f5242528411b262be447e61e2eb10f.… Which step in Figure 1 maps to the content of Figure 10? ⟳ Deletion/duplication/re… Deletion/duplication/re… defective_question
indonesiamobilemarketresearch-ag-… Which group accounts for the second largest share in te… ⟳ Christians Christians defective_question
reportq32015-151009093138-lva1-ap… As of Q3 2015, is vietnam's adoption rate of iOS 7 high… ⟳ Not answerable Not answerable defective_question
2303.08559v2.pdf What is the performance of filter-then-rerank methods (… ⟳ 72.3% 72.3% ambiguous_question
2307.09288v2.pdf list the top-3 models in Figure 3 ⟳ ['Vicuna13b-v1.3', 'PaL… ['Vicuna 33b-v1.3', 'Pa… ambiguous_question
2312.04350v3.pdf Which model performs the best on Cladder? ⟳ GPT-4 GPT-4 ambiguous_question
8e7c4cb542ad160f80fb3d795ada35d8.… What is the residential capacity of Staten Island from … ⟳ 435000000 435000000 ambiguous_question
AMAZON_2017_10K.pdf what is Amazon's FY2017 Operating Profit Margin Before … ⟳ 0.073 0.073 ambiguous_question
AMAZON_2017_10K.pdf what is Amazon's FY2017 debt to ebitda ratio? round you… ⟳ 1.93 1.93 ambiguous_question
BESTBUY_2023_10K.pdf what is invested capital of Best Buy for for the fiscal… ⟳ 13929 13929 ambiguous_question
COSTCO_2021_10K.pdf What does Costco rely heavily on for its financial perf… ⟳ the financial performan… the financial performan… ambiguous_question
COSTCO_2021_10K.pdf what is long-term debt of Costco in FY 2021? Anwser in … ⟳ 10314 10314 ambiguous_question
COSTCO_2021_10K.pdf what is total debt of COSTCO in FY 2021?Answer in milli… ⟳ 11407 11407 ambiguous_question
COSTCO_2021_10K.pdf what is total debt to EBITDA ratio of COSTCO in FY2021?… ⟳ 1.344 1.344 ambiguous_question
COSTCO_2021_10K.pdf what is Long-term Debt to Total Liabilities for COSTCO … ⟳ 0.25 0.25 ambiguous_question
COSTCO_2021_10K.pdf what is total debt to total assets for costco in FY 202… ⟳ 0.192 0.192 ambiguous_question
COSTCO_2021_10K.pdf What is common equity for COSTCO in FY2021? ⟳ 18078 18078 ambiguous_question
ISEP_student_handbook_2020.pdf Which compulsory ISEP courses does the students must ha… ⟳ ['GS5002', 'GS6001', 'G… ['GS5002', 'GS6001', 'G… ambiguous_question
Pew-Research-Center_Hispanic-Iden… What's the average value of all orange bars in the char… ⟳ 21 21 ambiguous_question
chapter8-geneticscompatibilitymod… In the case presented in Chapter 9, what color are the … ⟳ Not answerable Not answerable ambiguous_question
mmdetection-readthedocs-io-en-v2.… If I want to use the detector in the paper `SOLO: Segme… ⟳ DecoupledSOLOHead DecoupledSOLOHead ambiguous_question
welcome-to-nus.pdf How many people with white shirt are there in the Page … ⟳ Not answerable 1 ambiguous_question
2311.16502v3.pdf How many tables are included in Pages 105-110? Not answerable 1 should_be_answerable
Independents-Report.pdf From this report, among Clinton, G.W.Bush, and Obama, w… Not answerable None should_be_answerable
NUS-Business-School-BBA-Brochure-… From 2022 graduate employment survey, do graduates with… Not answerable No should_be_answerable
NUS-FASS-Graduate-Guidebook-2021-… Which of the following department does not provide a MB… Not answerable None should_be_answerable
PP_2021.04.22_voting-access_REPOR… Compared to October 2018, the proportion of Democrats w… Not answerable No should_be_answerable
PRE_2022.09.29_NSL-politics_REPOR… What proportion of the Spanish dominant Latinos express… Not answerable 46% should_be_answerable
afe620b9beac86c1027b96d31d396407.… How much higher was the dividend paid in 2003 compared … Not answerable 0.70 should_be_answerable
camry_ebrochure.pdf What models of wheel are introduced on pages 10 and 11? Not answerable ['18-in. black machined… should_be_answerable
csewt7zsecmmbzjufbyx-signature-24… From 2009 to 2013, as for the adviser's organic growth … Not answerable 0 should_be_answerable
csewt7zsecmmbzjufbyx-signature-24… How many years have there been more than 4,500 births (… Not answerable 0 should_be_answerable
e79deb02a0c0e87511080836c5d4347b.… What are the top2 texts of the yellow words in the docu… Not answerable ['Strategic Planning', … should_be_answerable
earthlinkweb-150213112111-convers… What is the sum of percentage of customers and employer… Not answerable 122 should_be_answerable
nielsen2015musicbizpresentation-f… What is the difference in total volume between the rank… Not answerable 1175000 should_be_answerable
owners-manual-2170416.pdf In the two styles shown in the "Parts and Features" sec… Not answerable No should_be_answerable
379f44022bb27aa53efd5d322c7b57bf.… what is the number of red logos in page 10? Not answerable Not answerable or 0 or … incomplete_answer
7c3f6204b3241f142f0f8eb8e1fefe7a.… What types of charts are in the document? Not answerable Not answerable or 0 or … incomplete_answer
8e7c4cb542ad160f80fb3d795ada35d8.… Which area of New York has more than 23% land area rezo… Not answerable Not answerable or 0 or … incomplete_answer
8e7c4cb542ad160f80fb3d795ada35d8.… What is the color of the bike in the picture on page 2? Not answerable Not answerable or 0 or … incomplete_answer
936c0e2c2e6c8e0c07c51bfaf7fd0a83.… What service specifications are associated with the SRM… ['Microsoft Oracle Open… ['Microsoft Oracle Open… incomplete_answer
GPL-Graduate-Studies-Professional… How many people with sun glassess are there in Page 5? Not answerable Not answerable or 0 or … incomplete_answer
Macbook_air.pdf List all the countries/regions mentioned in the "Regula… ['Canada', 'Europe', 'K… ['Canada', 'Europe', 'K… incomplete_answer
User_Manual_1500S_Classic_EN.pdf In the picture on page 50, what time does the red bar s… Not answerable Not answerable or 0 or … incomplete_answer
a5879805d70c854ea4361e43a84e3bb2.… what is the texts of the underlined italic words in pag… Not answerable Not answerable or 0 or … incomplete_answer
chapter8-geneticscompatibilitymod… What plants is on the cover of each chapter? Not answerable Not answerable or 0 or … incomplete_answer
ddoseattle-150627210357-lva1-app6… Which Youtube does the slides use to show the consequce… Not answerable Not answerable or 0 or … incomplete_answer
e639029d16094ea71d964e2fb953952b.… What is the yellow color italic texts in page 9? Not answerable Not answerable or 0 or … incomplete_answer
e79deb02a0c0e87511080836c5d4347b.… What is the title of the diagram on page 10? Not answerable Not answerable or 0 or … incomplete_answer
earthlinkweb-150213112111-convers… What is the job of the contact person in the picture at… Not answerable Not answerable or 0 or … incomplete_answer
formwork-150318073913-conversion-… Which stages of casting a tunnel framework require a co… Not answerable Not answerable or 0 or … incomplete_answer
san-francisco-11-contents.pdf Name a restaurant between 45th Ave. and 44th Ave.. Not answerable Not answerable or 0 or … incomplete_answer
dr-vorapptchapter1emissionsources… What are the four concrect facts of global challenges? ['Increasing world popu… None corpus_defect
dr-vorapptchapter1emissionsources… What are the ten concrect consequences of global challe… Not answerable None corpus_defect
dr-vorapptchapter1emissionsources… Among the global challenges and requirements, which of … ['World Mobility', 'Wor… None corpus_defect
dr-vorapptchapter1emissionsources… Among the global challenges, which of them would not be… Not answerable None corpus_defect
dr-vorapptchapter1emissionsources… What four elements are exhausted in the combustion proc… ['Nitrogen', 'Water', '… None corpus_defect
dr-vorapptchapter1emissionsources… What are the top 3 sources according to the total emiss… ['Power Generation', 'V… None corpus_defect
dr-vorapptchapter1emissionsources… List all the PM health effects that increse by more tha… ['Cough', 'Phlegm', 'Ad… None corpus_defect
dr-vorapptchapter1emissionsources… For first two gases that causes greenhouse effect, list… ['13.7%', '0.1%'] None corpus_defect
dr-vorapptchapter1emissionsources… One kind of gas is the siginificant contributor to clim… ['USA', 'Russia', 'Sout… None corpus_defect
dr-vorapptchapter1emissionsources… How many slides includes at least one chart? 24 None corpus_defect
BESTBUY_2023_10K.pdf what is interest coverage ratio for AMCOR'FY 2020? roun… 51.286 Not answerable should_be_unanswerable
Macbook_air.pdf I'm a Macbook Air user in Mexico. According to this gui… 1-800-275-2273 Not answerable should_be_unanswerable

Considered and rejected

Questions that looked defective and turned out to be deliberate negative samples. Upstream pairs many items with a near-identical twin: one side uses the document's real terminology, page, year or entity and is answerable; the other changes one word and is labelled Not answerable. That is the point — it tests whether a system invents an answer for a premise the document does not contain. These are left exactly as upstream wrote them.

Document Answerable twin Twin labelled Not answerable
2310.05634v2.pdf "…map to both [NA] and a list of sub-graph knowledge?" "…map to either [NA] or a list…"
tacl_a_00660.pdf "Among the three correction strategies…" "Among the three generation strategies…"
t480_ug_en.pdf "Which country or region codes … Mainland China?" → SC "Which license codes … Mainland China?"
formwork-…_95.pdf "Which stages … require a heater?" → Stage 5 "Which stages … require a cooler?"
e79deb02…pdf "…governor as mentioned on the first page" → Rick Scott "…on the last page"
e79deb02…pdf "title of the diagram on page 9" "title of the diagram on page 10"
chapter8-genetics…pdf "What animal is on the cover of each chapter?" → leopard "What plants is on the cover…"
san-francisco-11-contents.pdf "Name a restaurant between 36th and 37th Ave." → Cassava "…between 45th and 44th Ave."
san-francisco-11-contents.pdf "…most central part of San Francisco" → 178 "…most central part of Oakland"
PIP_Seniors…pdf "…tracking survey" dated 2013 the same question dated 2020 / 2022
ddoseattle…pdf "Which Youtube … blindly following data" "Which Youtube … weak data leadership"

The same shape guards the corpus elsewhere: several items name a company the document never discusses — Apple in a Godfrey Phillips annual report, Amazon in a Facebook/Twitter deck, Boeing in the Germanwings analysis, Amcor in a Best Buy 10-K — and are all labelled Not answerable.

Before rewording any question, check for an answerable twin. A Not answerable key with empty evidence_pages is very likely intentional; one with a concrete value and a cited page is where an annotator did the work and mislabelled the prompt. Twin structure shows what the annotator was trying to test — it does not guarantee that both sides are labelled correctly. Two of the corrections above sit on one side of a pair whose other side is untouched.

Two traps worth naming

Stating a convention can introduce a new one. Several finance ratios were unanswerable as written because the filing supports more than one reading of "long-term debt" or "before depreciation". Adding "taking total debt as all debt plus all lease liabilities" fixed the lease ambiguity and immediately created another: $41 of short-term borrowings disclosed only in Costco's Note 5 is literally "all debt", so a careful system included it and diverged again. The wording now names the balance-sheet lines. Prefer the most specific phrasing available; universal quantifiers move the boundary rather than removing it.

Percentages need their own base. Three answers were a survey percentage multiplied by the wrong denominator — the full sample instead of the subgroup the chart describes. A chart headed "% of internet users" applied to every respondent produces a number larger than the population that could possibly qualify. Where a count is derived from a percentage, the note records which base was used.

Caveat

These come from auditing questions that one system answered confidently and therefore scored zero — a biased slice, and one where annotation errors naturally concentrate. The corpus-wide error rate is unmeasured, and a broader sample would establish it. Coverage is also uneven by task type: the lookup, derive and compare slices have been audited end to end, count and enumerate have not.

Thirteen entries from the first review pass record the change but not the reasoning; their note says so explicitly. They were checked against the cited pages at the time, but the derivation was not written down and has not been reconstructed.

Disagreements are useful. Open an issue naming the entry and the page you read; some questions turn out to be defective rather than mislabelled, which is why defective_question exists as a category.