R15.03.01.04 · 任务小类

基于图像的原文转写与翻译

Image-based source transcription & translation

Transcribe text from scans, manuscripts or screenshots and translate it with cited glyph readings.

20 条本项目编写的提案 · 分类 v1.1 · 新增基准运行 0 次。下方验证方式与部分得分规则是设计，不是执行结果。

历史小类地址：R10.07.01.04 · 提案 ID 保留原值。

典型应用: image viewers · OCR tools · text editors
小类 GUI 需求标签: G2 · GUI 必要

任务提案 · 20

展开查看完整原文与建议检查方式；两种界面均保留提案英文原文。

R10.07.01.04-T01 Diplomatic transcription of a handwritten 1890s letter

原始指令

Transcribe the 2-page handwritten English letter in input/letter_1893_p1.jpg and _p2.jpg diplomatically into out/letter.txt: keep original line breaks and spelling, expand abbreviations in [square brackets], mark illegible words as [illeg.], and add the page break marker '--- p2 ---'.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 type · 文字与表单
工作产物: WP01 · 文档
条件: 未声明
提案难度: 3 / 5

建议验证方式

CER <= 5 pct vs gold diplomatic transcription after normalizing whitespace; line-break alignment >= 0.9; abbreviation expansions bracketed (6 gold cases); page marker present. OCR engines are absent from the agent VM and cloud OCR endpoints are firewalled, so reading must happen by zooming the scans.

建议部分得分

55% character accuracy; 20% line structure; 15% abbreviation handling; 10% page marker and illegible marks

R10.07.01.04-T02 Identify a Yi manuscript glyph from a phonetic gloss

原始指令

input/yi_folio_12.jpg is a classical Yi manuscript page; input/gloss.txt gives a phonetic gloss for one glyph in line 4. Find that glyph, save its bounding box as out/bounding_box.json [x, y, w, h] in image pixels, and write out/report.md with the transcription of line 4 in Yi script, a word-by-word gloss and an English translation citing each glyph reading.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · text editor · Unicode input
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 geom · 精确几何 type · 文字与表单
工作产物: WP01 · 文档
条件: 未声明
提案难度: 5 / 5

建议验证方式

bbox IoU >= 0.5 with the gold glyph box; line 4 transcription CER <= 10 pct vs gold (Yi Unicode); translation scored against expert reference by chrF >= 45 plus LLM rubric for citation completeness (only for that milestone). OCR and MT endpoints firewalled; no Yi OCR tool exists on the VM.

建议部分得分

35% correct glyph located; 30% line transcription; 20% translation; 15% glyph citations (rubric)

R10.07.01.04-T03 Transcribe and translate a Japanese app UI screenshot

原始指令

input/settings_ja.png is a Japanese mobile settings screen with 12 strings. Transcribe each string and translate it into English in out/strings.xliff (XLIFF 1.2), one trans-unit per string with id 's01'..'s12' ordered top to bottom, and the source string's pixel bbox in a note.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: vis · 视觉判断 type · 文字与表单 geom · 精确几何
工作产物: WP02 · 记录 / 表格
条件: 未声明
提案难度: 3 / 5

建议验证方式

XLIFF validates; 12 units in visual order; source exact-match accuracy >= 11/12 vs gold (NFC); bbox IoU >= 0.5 per string; translations checked against glossary terms (UI terms in input/glossary.csv) and chrF >= 55. OCR binaries absent and cloud OCR/MT firewalled.

建议部分得分

40% source transcription; 20% order and bboxes; 30% translations; 10% valid XLIFF

R10.07.01.04-T04 Latin inscription in Leiden conventions with translation

原始指令

Transcribe the Roman funerary inscription in input/stele_photo.jpg (raking light, weathered) into out/edition.txt using Leiden conventions (abbreviations expanded in parentheses, restorations in square brackets, line numbers), then provide an English translation in out/translation.txt.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 type · 文字与表单
工作产物: WP01 · 文档
条件: 未声明
提案难度: 4 / 5

建议验证方式

Edition compared to gold: letter-level CER on preserved letters <= 8 pct; expansion and restoration markup F1 >= 0.8; line numbering exact. Translation chrF >= 50 vs reference and key name/age terms present. OCR engines absent; cloud services firewalled.

建议部分得分

40% preserved letters; 25% Leiden markup; 10% line numbers; 25% translation

R10.07.01.04-T05 Fraktur newspaper clipping to text and English

原始指令

input/zeitung_1912.tif is a scanned German Fraktur newspaper article (one column, 38 lines). Transcribe it into out/de.txt preserving original spelling (long s as 's', ligatures resolved), then translate it into English in out/en.txt paragraph by paragraph.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · OCR tools · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: vis · 视觉判断 type · 文字与表单 code · 代码与命令
工作产物: WP01 · 文档
条件: 未声明
提案难度: 3 / 5

建议验证方式

German CER <= 3 pct vs gold; paragraph count equal; English chrF >= 55 vs reference translation and the 5 named entities preserved. OCR (e.g. Tesseract Fraktur model) allowed but must be corrected (hybrid).

建议部分得分

50% German transcription accuracy; 15% paragraph structure; 35% translation

R10.07.01.04-T06 Classical Chinese woodblock page in reading order

原始指令

Transcribe the woodblock-printed page input/woodblock_p7.jpg (vertical, right-to-left, with interlinear commentary in small characters) into out/main.txt (main text) and out/commentary.txt, normalizing variant characters to standard traditional forms and listing each normalization in out/variants.csv. Translate the first two passages into English.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 type · 文字与表单
工作产物: WP01 · 文档
条件: 未声明
提案难度: 5 / 5

建议验证方式

Main text CER <= 5 pct and commentary CER <= 10 pct vs gold; reading order correct (column sequence); variants.csv F1 >= 0.8 vs gold list; translation chrF >= 40 plus LLM rubric for meaning (that milestone only). OCR engines absent and online OCR firewalled.

建议部分得分

35% main text; 20% commentary separated; 20% variant normalization; 25% translation

R10.07.01.04-T07 Extract fields from a handwritten Arabic receipt

原始指令

From input/receipt_ar.jpg (handwritten Arabic shop receipt) extract into out/receipt.json: date as written (Hijri) and converted to Gregorian ISO, shop name transliterated (ALA-LC) and translated, buyer name, each line item (description translated, quantity, unit price) and the total.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: vis · 视觉判断 type · 文字与表单 nav · 平移 / 缩放 / 滚动
工作产物: WP02 · 记录 / 表格
条件: 未声明
提案难度: 3 / 5

建议验证方式

JSON schema valid; Hijri date exact and Gregorian conversion correct; numbers exact (Eastern Arabic numerals read correctly); item descriptions match gold semantics (term list); total equals sum and gold. OCR binaries absent; cloud OCR/MT firewalled.

建议部分得分

25% dates; 40% line items and numbers; 20% names and transliteration; 15% total

R10.07.01.04-T08 Romanize street signs from photos

原始指令

input/signs/ has 10 street-sign photos (5 Korean, 5 Thai). For each, write the sign text exactly in the original script and its romanization (Revised Romanization of Korean; RTGS for Thai) to out/signs.csv with file, script_text, romanization.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: vis · 视觉判断 type · 文字与表单
工作产物: WP02 · 记录 / 表格
条件: 未声明
提案难度: 2 / 5

建议验证方式

Exact-match accuracy of script_text >= 9/10 (NFC); romanization exact per standard for >= 8/10 (gold). OCR binaries absent from the VM and online OCR/translate endpoints firewalled.

建议部分得分

55% original-script text; 45% romanization

R10.07.01.04-T09 Spanish colonial ledger page to a structured table

原始指令

Transcribe the 1780s Spanish account ledger page input/libro_cuentas_f23.jpg (columns: date, concept, pesos, reales) into out/ledger.csv with original spelling, add a 'concept_en' column translated into English, and check the page total against the written sum, noting any discrepancy in out/notes.txt.

平台: mac · macOS win · Windows
应用: image viewer · spreadsheet · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 type · 文字与表单
工作产物: WP02 · 记录 / 表格
条件: multiapp · 多应用
提案难度: 4 / 5

建议验证方式

Rows equal gold count (27); numeric columns exact for >= 25 rows; concept CER <= 8 pct; translations term-checked; notes identify the seeded 2-real addition error. OCR engines absent and cloud OCR firewalled.

建议部分得分

35% numeric values; 25% concept transcription; 20% translations; 20% discrepancy found

R10.07.01.04-T10 Translate a French menu photo keeping prices

原始指令

Translate the photographed French bistro menu input/menu_fr.jpg into English as out/menu_en.md, keeping section headings, every dish and its exact price in euros, and adding the original French dish name in italics after each translation.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: vis · 视觉判断 type · 文字与表单
工作产物: WP01 · 文档
条件: 未声明
提案难度: 1 / 5

建议验证方式

All 24 dishes present with exact prices; French names exact (accent-sensitive) for >= 22; section structure equal; translation term accuracy against gold glossary >= 0.85. OCR binaries absent and online translate/OCR firewalled.

建议部分得分

35% dishes and prices; 25% French names; 30% translations; 10% structure

R10.07.01.04-T11 Place-name labels from an 18th-century map with locations

原始指令

From the map scan input/map_1750_sheet3.tif (12000x9000 px), transcribe every settlement label inside the red-boxed region (34 labels), record each label's bbox, and give the modern name where identifiable, into out/places.json [{text, bbox, modern_name}].

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 geom · 精确几何 type · 文字与表单
工作产物: WP10 · 标注
条件: long · 长流程
提案难度: 4 / 5

建议验证方式

Labels matched to gold by bbox IoU >= 0.4: recall >= 0.9; text CER <= 10 pct on matched labels; modern names exact for >= 0.8 of the 20 identifiable. OCR engines are not installed on the VM; the map requires zoom/pan inspection.

建议部分得分

35% label recall with boxes; 35% transcription accuracy; 30% modern identifications

R10.07.01.04-T12 Exact text of a low-res Russian error dialog

原始指令

input/error_ru.png is a low-resolution screenshot of a Russian error dialog from accounting software. Write out/error.json with the exact title, message and button labels in Russian, an English translation, and the product name and version shown in the dialog.

平台: win · Windows mac · macOS
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: vis · 视觉判断 type · 文字与表单 nav · 平移 / 缩放 / 滚动
工作产物: WP12 · 答案 / 决策
条件: 未声明
提案难度: 2 / 5

建议验证方式

Russian strings exact-match vs gold (NFC, punctuation-sensitive) for title, message and 3 buttons; version string exact; translation chrF >= 55. OCR binaries absent; cloud OCR firewalled.

建议部分得分

50% Russian text exact; 20% product and version; 30% translation

R10.07.01.04-T13 Damaged Sanskrit folio: IAST transliteration with lacunae

原始指令

Transliterate the damaged Devanagari manuscript folio input/folio_42r.jpg into IAST in out/iast.txt, line by line, marking lacunae as [...] with the estimated number of akṣaras (e.g. [...3...]) and uncertain readings with '?'. Then give out/akshara_boxes.json with bboxes for the 5 uncertain akṣaras.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 type · 文字与表单 geom · 精确几何
工作产物: WP01 · 文档
条件: 未声明
提案难度: 5 / 5

建议验证方式

IAST CER <= 6 pct on legible text vs gold; lacuna positions and length estimates within +/-1 akṣara for >= 4/5; uncertain boxes IoU >= 0.4 with gold. OCR engines absent from the VM; online OCR firewalled.

建议部分得分

45% transliteration accuracy; 25% lacunae marked; 30% uncertain akṣaras located

R10.07.01.04-T14 Patent drawing reference numerals and legend translation

原始指令

For the Japanese patent drawing sheet input/jp_patent_fig3.png, list every reference numeral with its leader target described in English, transcribe the figure legend (図3...) and translate it, using the numeral descriptions in input/claims_ja.txt. Output out/fig3.json.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: vis · 视觉判断 nav · 平移 / 缩放 / 滚动 type · 文字与表单
工作产物: WP02 · 记录 / 表格
条件: 未声明
提案难度: 3 / 5

建议验证方式

Numeral set equals gold (F1 >= 0.95); each numeral's English part name matches gold term list >= 0.85; legend exact-match Japanese and translation chrF >= 50. OCR binaries absent and online OCR/MT firewalled.

建议部分得分

35% numerals found; 35% part names; 30% legend and translation

R10.07.01.04-T15 Hebrew gravestone inscription with calendar conversion

原始指令

Transcribe the Hebrew inscription on input/matzevah.jpg (weathered, with abbreviations such as פ"נ and תנצב"ה) into out/inscription.txt, expand abbreviations, translate into English, and convert the Hebrew death date (in letters) to the Gregorian date in out/date.txt (ISO).

平台: mac · macOS win · Windows
应用: image viewer · text editor · calendar converter
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: vis · 视觉判断 type · 文字与表单 nav · 平移 / 缩放 / 滚动
工作产物: WP01 · 文档
条件: 未声明
提案难度: 4 / 5

建议验证方式

Hebrew transcription CER <= 6 pct vs gold; abbreviation expansions exact for the 4 gold cases; Gregorian date exact (grader converts gold Hebrew date); translation names and relations correct (term check). OCR engines absent; cloud OCR firewalled.

建议部分得分

40% transcription; 15% abbreviations; 25% date conversion; 20% translation

R10.07.01.04-T16 Manga page balloons in reading order with translations

原始指令

For the manga page input/manga_p14.png, locate every speech balloon and caption (14), give each a bbox, transcribe the Japanese text, and translate it into English that fits naturally, ordered by manga reading order (right-to-left, top-to-bottom). Output out/balloons.json.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: vis · 视觉判断 geom · 精确几何 type · 文字与表单
工作产物: WP10 · 标注
条件: 未声明
提案难度: 3 / 5

建议验证方式

Balloon recall >= 13/14 at IoU >= 0.5; order matches gold (Kendall tau >= 0.9); Japanese CER <= 8 pct; translations chrF >= 40 vs reference plus LLM rubric on naturalness (that milestone only). OCR binaries absent; cloud OCR/MT firewalled.

建议部分得分

30% balloons located; 15% reading order; 30% transcription; 25% translation

R10.07.01.04-T17 Ottoman Turkish document with cited uncertain readings

原始指令

Transliterate the Ottoman Turkish (Arabic-script) court record input/sicil_1873.jpg into modern Turkish Latin transliteration (IJMES style) in out/translit.txt, translate it into English, and for the 5 most uncertain words give bbox, your reading, and an alternative reading in out/uncertain.json.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 type · 文字与表单 geom · 精确几何
工作产物: WP01 · 文档
条件: 未声明
提案难度: 5 / 5

建议验证方式

Transliteration word accuracy >= 0.85 vs expert gold; translation chrF >= 45 and key legal terms present; uncertain words: bbox IoU >= 0.4 with gold-flagged hard words for >= 3/5. OCR engines absent; online OCR/MT firewalled.

建议部分得分

40% transliteration; 30% translation; 30% uncertain readings cited

R10.07.01.04-T18 Typewritten Italian memo with stamps and marginalia

原始指令

Transcribe the typewritten 1943 Italian memo input/memo_1943.jpg into out/body.txt (main text) and out/annotations.json (each stamp and handwritten marginal note with text, type and bbox), then translate the body into English in out/body_en.txt.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · OCR tools · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: vis · 视觉判断 type · 文字与表单 code · 代码与命令
工作产物: WP01 · 文档
条件: 未声明
提案难度: 3 / 5

建议验证方式

Body CER <= 2 pct; annotations: 6 gold items matched by IoU >= 0.4 with type correct, handwritten note text CER <= 15 pct; translation chrF >= 55. OCR permitted for the typed body (hybrid); stamps and handwriting require visual reading.

建议部分得分

35% body transcription; 35% annotations located and read; 30% translation

R10.07.01.04-T19 Bilingual transcript from overlapping chat screenshots

原始指令

input/chat/ holds 6 overlapping screenshots of a WeChat conversation in Chinese. Reconstruct the full conversation once (no duplicated messages from overlaps) into out/chat.csv with speaker, timestamp, original text, English translation, in chronological order.

平台: mac · macOS win · Windows
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: vis · 视觉判断 type · 文字与表单 multi · 跨应用
工作产物: WP02 · 记录 / 表格
条件: 未声明
提案难度: 3 / 5

建议验证方式

Message count equals gold 41 (duplicates penalized); speaker and timestamp exact for >= 39; Chinese CER <= 4 pct; translations chrF >= 50. OCR binaries absent and online OCR/MT firewalled.

建议部分得分

30% de-duplicated message set; 20% speakers and timestamps; 30% transcription; 20% translation

R10.07.01.04-T20 Middle English charter with abbreviation glyph glossary

原始指令

Transcribe the Middle English charter input/charter_1420.jpg semi-diplomatically into out/charter.txt (abbreviations expanded in italics markup _..._), and build out/abbrev.json listing 10 distinct abbreviation marks you expanded, each with an example bbox, the mark's name (e.g. macron, 'per' bar) and its expansion.

平台: mac · macOS win · Windows lin · Linux
应用: image viewer · text editor
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: nav · 平移 / 缩放 / 滚动 vis · 视觉判断 geom · 精确几何 type · 文字与表单
工作产物: WP01 · 文档
条件: 未声明
提案难度: 4 / 5

建议验证方式

Expanded-text CER <= 6 pct vs gold; expansion markup positions F1 >= 0.8; abbrev.json: 10 entries whose bboxes hit gold abbreviation instances (IoU >= 0.3) with correct mark type and expansion for >= 8. OCR engines absent; cloud OCR firewalled.

建议部分得分

45% transcription; 20% expansion markup; 35% abbreviation glossary with boxes

已有研究映射

以下沿用研究目录中的基准与小类映射，不是逐题标注、执行记录或产品背书。

Agents' Last Exam (ALE)

基于图像的原文转写与翻译

任务提案 · 20

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

已有研究映射

同工作流族的小类