I joined this competition late, and up until three days before the submission deadline I was mostly focused on cleaning up and curating train.csv.\nBecause of that, I didnât have enough time to properly verify local CV, and in the end I had to rely on LB score and gut feeling, which is something I regret.
train.csvThe original train.csv had 1,561 records, but after doing the following work I expanded it to 2,600 records in the end:
train.csv (I found 27 missing items).train.csv described below).I also fixed issues like the following in train.csv:
train.csv (this was relatively easy to recover by comparing with published_texts.csv).<gap> tokens that splitting the transliteration and translation was not feasible, since they were likely harmful for model training (surprisingly, I found only 3 such records).I also found issues like the following in many places, but since the volume was too large and I suspected that similar issues existed in test.csv used for the final evaluation, I intentionally left them as-is:
For fixing train.csv using the PDF, I used a very primitive method: taking screenshots of the relevant parts and having ChatGPT transcribe them.
For transcription
\n> NOTE: The original version of this prompt was written in Japanese. The text below is an English translation.\n\n# Purpose (request to you)\nThe **screenshot images** Iâm going to send contain **Akkadian transliteration** on the left (or top) and the corresponding **English translation** on the right (or bottom). \nPlease **transcribe both accurately** and return them in a format that is **easy to copy & paste into a spreadsheet**.\n\nIn some cases, a single screenshot might not be enough and the content may be split across multiple screenshots. In that case, please infer the correct order of the screenshots and transcribe accordingly.\n\n---\n\n## Response format (copy/paste first)\nIn principle, please reply every time with the following 2-line set (no extra explanation needed):\n\ntransliteration (Akkadian text) <newline>\n<separator>\ntranslation (English text)\n\n- `<newline>` means a line break in the ChatGPT UI; `<separator>` also means a line break in the ChatGPT UI.\n- Keep the Akkadian text and the English text to **one line each**.\n\n---\n\n## Handling body text vs. notes (important)\n- The image may contain lines other than the âbody textâ (e.g., notes, headings, layout notes, page numbers, line numbers), such as `Seal A`, `rev.`, `obv.`, `(remainder obv. missing)`, `(3.5 cm un-inscribed)`, etc.\n- Please **extract only the body text** and **exclude note-like lines** from the transcription output.\n\nExample output (when NOTE-like lines are present):\nAkkadian: transcribed transliteration (body) <newline>\n<newline>\nEnglish: transcribed translation (body)\n\n---\n\n## Transcription rules (important)\n### 1) Normalization / replacements (required)\nFor expressions indicating missing/damaged/illegible text, if the meaning is the same, **unify everything to `<gap>`**.\n\n- `x` â `<gap>`\n- `[x]` â `<gap>`\n- `âŚ` (ellipsis) â `<gap>`\n- `(break)` â `<gap>`\n- `(large break)` â `<gap>`\n- `(n broken lines)` â `<gap>` (itâs OK to use `<gap>` even when `n` is a number)\n- If `<gap>` appears multiple times consecutively, merge into one; i.e., `<gap> <gap>` â `<gap>`\n- **Any other âmissing/broken/unreadableâ indicators** â `<gap>`\n\nFor determinatives, normalize as follows:\n\n- `(d)` â `{d}`\n- `(ki)` â `{ki}`\n\nFor parenthesized words:\n\n- `(TĂG)` â `TĂG` (remove the parentheses)\n\n---\n\n## 2) Things to keep as-is\n- Keep uppercase/lowercase, apostrophes, hyphens, subscript numerals (e.g., `â`), special characters (e.g., `ĹĄ ᚣ ᚠḍ`), accents, etc. **exactly as shown**\n- Keep symbols/separators (like `-` and `.`) **as in the image**, except where the replacement rules apply\n- `<gap>` must always be exactly `<gap>` in **lowercase**\n\n---\n\n## 3) Uncertainty\n- If there is a character/word you truly cannot identify, you may replace that part with `<gap>`\n- However, preserve whatever you can read (donât turn everything into `<gap>`)\n\n---\n\n## Minimal competition context (for clarity)\n- Input: Akkadian (cuneiform) transliteration = transliteration\n- Output: its English translation = translation\n- The goal is to create transliteration â English pairs, so please extract them as **two columns of text**\n\n---\n\n## Requests to make the work faster and more accurate\n- If possible, extract the âtransliteration partâ and the âEnglish translation partâ separately so they **donât get mixed**\n- Since punctuation and missing closing parentheses are common errors, do a quick check that the left/right (or top/bottom) alignment hasnât drifted before you output\n- The output is intended to be pasted into a spreadsheet, so please **avoid extra decoration** (bullets, block quotes, etc.)\nFor splitting long texts with alignment preserved
\n> NOTE: The original version of this prompt was written in Japanese. The text below is an English translation.\n\n# Purpose (request to you)\n\nThe text and screenshot images Iâm going to send are candidate training data for the Kaggle competition **Deep Past Challenge - Translate Akkadian to English**. \nFor a single long record, please split it into **a small number (around 2â5) of independent records**, while being careful not to break the alignment between the **Akkadian transliteration** and the **English translation**.\n\nThe goal is to make each record easier to handle with a ByT5 model by keeping it **under 1024 bytes**. \nHowever, if splitting would likely break the alignment, it is better to return it as âcannot splitâ than to over-split.\n\n---\n\n## Most important policy (v2: donât over-split)\n\n- Target number of splits: **2â5** (prefer **2â4** if possible)\n- Split only when necessary\n - Add extra split boundaries only when they are truly needed to keep chunks under 1024 bytes or to preserve alignment\n- Avoid producing many very short chunks\n - If you canât count bytes exactly, err on the safe side, but **donât mass-produce tiny chunks**\n - As a rough guide, aim for **large chunks (roughly 700â1000 bytes)** when possible\n- If it still wonât fit within 5 splits or alignment would break, return `needs_manual_review` (do not force 6+ splits)\n\n---\n\n## Background (minimal)\n\n- In this competition, we train a model to generate English translations from Akkadian transliterations.\n- ByT5 is byte-level, so **UTF-8 byte length** matters more than character count.\n- On the source side, we usually add the prefix `translate Akkadian to English: `.\n- In practice, we want both source and target to be **under 1024 bytes** per record.\n- You donât need to compute byte length perfectly every time. If unsure, make it a bit shorter for safety.\n\n---\n\n## What I want you to do\n\n1. Read the given `transliteration` and `translation`\n2. Refer to the screenshot images if needed\n3. Find only a small number of boundaries where the **semantic alignment is clear** (donât add too many)\n4. Split the record into **2â5** `transliteration_chunk` / `translation_chunk` pairs (choose the minimum split that works)\n5. If splitting is risky, return `needs_manual_review`\n\n---\n\n## Absolute rules\n\n### 1) This is âsplittingâ, not âeditingâ\n\n- For the body text, use the input text as-is and, in principle, only take **substrings** from the provided `transliteration` / `translation`\n- Do not fill in missing text, paraphrase, summarize, or re-translate\n- Even if you notice OCR mistakes or unnatural English, **do not fix them in this task**\n\n### 2) Preserve the order\n\n- When you stack the chunks from top to bottom, they must preserve the original order of the `transliteration` / `translation`\n\n### 3) Only cut at boundaries visible on both sides\n\n- Sentence/clause boundaries on the English side\n- Boundaries between itemized lists\n- Breaks in letter templates / formulaic phrases\n- Breaks between imperative/conditional sections\n- Clear line-group boundaries on the transliteration side\n\nOnly cut at places where a coherent âunitâ is visible on both sides.\n\n### 4) Do not cut in the middle of these\n\n- Numbers, fractions, decimals, tax rates\n- Amount expressions like shekel / mina groupings\n- Personal names, divine names, place names\n- Phrases like `son of ...` or `belonging to ...`\n- Inside quotation marks\n- Inside an item name\n\n### 5) If itâs not safe, hold it\n\nIn the following cases, do not force a split; return `needs_manual_review`:\n\n- Even with the images, the aligned boundary cannot be determined\n- The English translation in train looks broken in the middle\n- The source/target are misaligned and 1:1 correspondence is unclear\n- Keeping it within 2â5 splits would inevitably break numeric/proper-noun sequences\n\n---\n\n## How to use the images\n\nUse the screenshots to decide **where to cut so that transliteration and translation stay aligned**.\n\nImportant:\n\n- The Akkadian and English text in the PDF may not exactly match the text provided in the chat.\n- This may be because the chat text has already been normalized (e.g., `<gap>` replacements, symbol cleanup, etc.).\n- Therefore, use the PDF as a reference to find alignment boundaries, but for the final output body text, **prefer the chat-provided text**.\n- Do not rewrite the body text to match the PDFâs visual appearance.\n\nPriority when using images:\n\n1. Verify alignment boundaries between `transliteration` and `translation`\n2. Resolve ambiguous boundaries that are hard to see from text alone\n3. Decide what spans form a coherent aligned unit\n\nDo not create new translations even if you use the images.\n\n---\n\n## Boundary heuristics (for a small number of splits)\n\n### Good candidates (high priority)\n\n- After an introductory formula like `a-na ... qĂ-bi-ma um-ma ...` (but donât split into a tiny chunk just for the intro)\n- English expressions that indicate a section change, e.g. `all this ...`, `thereof ...`, `therefore ...`, `if ... then ...`\n- Places where a large block of an item list changes (avoid splitting each small item)\n- Where repetition of the same pattern ends as a block\n\n### Tempting but usually forbidden\n\n- Splitting every English sentence (often creates too many chunks)\n- Splitting every small list item (avoid)\n\n---\n\n## Output format (strict)\n\nReturn **only a JSON array**. No extra explanation text.\n\n### When splitting is possible (2â5 chunks)\n\n```json\n[\n {\n \"record_id\": \"original record id\",\n \"split_id\": \"original record id__01\",\n \"status\": \"ok\",\n \"transliteration_chunk\": \"split transliteration chunk (must be a substring of the input)\",\n \"translation_chunk\": \"corresponding translation chunk (must be a substring of the input)\",\n \"boundary_reason\": \"one sentence describing why you cut here (alignment rationale)\",\n \"confidence\": \"high\"\n }\n]\n```\n\n### When splitting is risky\n\n```json\n[\n {\n \"record_id\": \"original record id\",\n \"split_id\": \"original record id__00\",\n \"status\": \"needs_manual_review\",\n \"transliteration_chunk\": \"\",\n \"translation_chunk\": \"\",\n \"boundary_reason\": \"cannot cut into 2â5 chunks without breaking alignment\",\n \"confidence\": \"low\"\n }\n]\n```\n\n---\n\n## `confidence` guide\n\n- `high`: boundaries are very clear from both images and text\n- `medium`: mostly reasonable, but alignment is slightly weak on one side\n- `low`: the original train text may be missing/broken/misaligned\n\n---\n\n## When you receive an actual request, always follow these\n\n- Return only a JSON array\n- Strongly prefer 2â5 splits and do not add too many boundaries\n- Remember this is splitting, not editing\n- Do not add words that are not in the provided body text, even if you look at images\n- If unsure, return `needs_manual_review`\nAt first I tried to clean up train.csv using Codex agents and the skills feature, but I couldnât control it well, and in the end I did it manually.
I started from the public notebook âLB 35.9 Ensembling & Post Processing Baselineâ, then increased the number of models and made small changes to preprocessing and postprocessing.\nThe last two submissions differed only in the number of training epochs; despite concerns about overfitting, the one trained for more epochs achieved a higher Private LB score.
\nBoth preprocessing and postprocessing are differences from âLB 35.9 Ensembling & Post Processing Baselineâ (only the differences are listed here).
\n_EXACT_FRAC_RE + _canon_decimal(float)):1/2 and 2 1/2 into Unicode vulgar fractions (e.g., ½)0.16666... â â
).5 values like 2.5 as-is (by design)Postprocessing and the final submission formatting are also differences from âLB 35.9 Ensembling & Post Processing Baselineâ (only the differences are listed here).
\n5/12 shekel(s): baseline â
shekel 15 grains â â
shekel 15 grains1/2 / 2 1/2 / decimals â Unicode fractions)you / she, keep the left side (you) and drop the right side (she)1/2\" / ') instead (baseline deletes them)() in bulk).. / xx..., etc.), and as a safeguard normalized ḍ/Ḫ to h/H<gap>validate_submission to check columns/row count/empty translations/duplicate ids (baseline mainly logs the number of empty outputs)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)I joined this competition late, and up until three days before the submission deadline I was mostly focused on cleaning up and curating train.csv.\nBecause of that, I didnât have enough time to properly verify local CV, and in the end I had to rely on LB score and gut feeling, which is something I regret.
train.csvThe original train.csv had 1,561 records, but after doing the following work I expanded it to 2,600 records in the end:
train.csv (I found 27 missing items).train.csv described below).I also fixed issues like the following in train.csv:
train.csv (this was relatively easy to recover by comparing with published_texts.csv).<gap> tokens that splitting the transliteration and translation was not feasible, since they were likely harmful for model training (surprisingly, I found only 3 such records).I also found issues like the following in many places, but since the volume was too large and I suspected that similar issues existed in test.csv used for the final evaluation, I intentionally left them as-is:
For fixing train.csv using the PDF, I used a very primitive method: taking screenshots of the relevant parts and having ChatGPT transcribe them.
For transcription
\n> NOTE: The original version of this prompt was written in Japanese. The text below is an English translation.\n\n# Purpose (request to you)\nThe **screenshot images** Iâm going to send contain **Akkadian transliteration** on the left (or top) and the corresponding **English translation** on the right (or bottom). \nPlease **transcribe both accurately** and return them in a format that is **easy to copy & paste into a spreadsheet**.\n\nIn some cases, a single screenshot might not be enough and the content may be split across multiple screenshots. In that case, please infer the correct order of the screenshots and transcribe accordingly.\n\n---\n\n## Response format (copy/paste first)\nIn principle, please reply every time with the following 2-line set (no extra explanation needed):\n\ntransliteration (Akkadian text) <newline>\n<separator>\ntranslation (English text)\n\n- `<newline>` means a line break in the ChatGPT UI; `<separator>` also means a line break in the ChatGPT UI.\n- Keep the Akkadian text and the English text to **one line each**.\n\n---\n\n## Handling body text vs. notes (important)\n- The image may contain lines other than the âbody textâ (e.g., notes, headings, layout notes, page numbers, line numbers), such as `Seal A`, `rev.`, `obv.`, `(remainder obv. missing)`, `(3.5 cm un-inscribed)`, etc.\n- Please **extract only the body text** and **exclude note-like lines** from the transcription output.\n\nExample output (when NOTE-like lines are present):\nAkkadian: transcribed transliteration (body) <newline>\n<newline>\nEnglish: transcribed translation (body)\n\n---\n\n## Transcription rules (important)\n### 1) Normalization / replacements (required)\nFor expressions indicating missing/damaged/illegible text, if the meaning is the same, **unify everything to `<gap>`**.\n\n- `x` â `<gap>`\n- `[x]` â `<gap>`\n- `âŚ` (ellipsis) â `<gap>`\n- `(break)` â `<gap>`\n- `(large break)` â `<gap>`\n- `(n broken lines)` â `<gap>` (itâs OK to use `<gap>` even when `n` is a number)\n- If `<gap>` appears multiple times consecutively, merge into one; i.e., `<gap> <gap>` â `<gap>`\n- **Any other âmissing/broken/unreadableâ indicators** â `<gap>`\n\nFor determinatives, normalize as follows:\n\n- `(d)` â `{d}`\n- `(ki)` â `{ki}`\n\nFor parenthesized words:\n\n- `(TĂG)` â `TĂG` (remove the parentheses)\n\n---\n\n## 2) Things to keep as-is\n- Keep uppercase/lowercase, apostrophes, hyphens, subscript numerals (e.g., `â`), special characters (e.g., `ĹĄ ᚣ ᚠḍ`), accents, etc. **exactly as shown**\n- Keep symbols/separators (like `-` and `.`) **as in the image**, except where the replacement rules apply\n- `<gap>` must always be exactly `<gap>` in **lowercase**\n\n---\n\n## 3) Uncertainty\n- If there is a character/word you truly cannot identify, you may replace that part with `<gap>`\n- However, preserve whatever you can read (donât turn everything into `<gap>`)\n\n---\n\n## Minimal competition context (for clarity)\n- Input: Akkadian (cuneiform) transliteration = transliteration\n- Output: its English translation = translation\n- The goal is to create transliteration â English pairs, so please extract them as **two columns of text**\n\n---\n\n## Requests to make the work faster and more accurate\n- If possible, extract the âtransliteration partâ and the âEnglish translation partâ separately so they **donât get mixed**\n- Since punctuation and missing closing parentheses are common errors, do a quick check that the left/right (or top/bottom) alignment hasnât drifted before you output\n- The output is intended to be pasted into a spreadsheet, so please **avoid extra decoration** (bullets, block quotes, etc.)\nFor splitting long texts with alignment preserved
\n> NOTE: The original version of this prompt was written in Japanese. The text below is an English translation.\n\n# Purpose (request to you)\n\nThe text and screenshot images Iâm going to send are candidate training data for the Kaggle competition **Deep Past Challenge - Translate Akkadian to English**. \nFor a single long record, please split it into **a small number (around 2â5) of independent records**, while being careful not to break the alignment between the **Akkadian transliteration** and the **English translation**.\n\nThe goal is to make each record easier to handle with a ByT5 model by keeping it **under 1024 bytes**. \nHowever, if splitting would likely break the alignment, it is better to return it as âcannot splitâ than to over-split.\n\n---\n\n## Most important policy (v2: donât over-split)\n\n- Target number of splits: **2â5** (prefer **2â4** if possible)\n- Split only when necessary\n - Add extra split boundaries only when they are truly needed to keep chunks under 1024 bytes or to preserve alignment\n- Avoid producing many very short chunks\n - If you canât count bytes exactly, err on the safe side, but **donât mass-produce tiny chunks**\n - As a rough guide, aim for **large chunks (roughly 700â1000 bytes)** when possible\n- If it still wonât fit within 5 splits or alignment would break, return `needs_manual_review` (do not force 6+ splits)\n\n---\n\n## Background (minimal)\n\n- In this competition, we train a model to generate English translations from Akkadian transliterations.\n- ByT5 is byte-level, so **UTF-8 byte length** matters more than character count.\n- On the source side, we usually add the prefix `translate Akkadian to English: `.\n- In practice, we want both source and target to be **under 1024 bytes** per record.\n- You donât need to compute byte length perfectly every time. If unsure, make it a bit shorter for safety.\n\n---\n\n## What I want you to do\n\n1. Read the given `transliteration` and `translation`\n2. Refer to the screenshot images if needed\n3. Find only a small number of boundaries where the **semantic alignment is clear** (donât add too many)\n4. Split the record into **2â5** `transliteration_chunk` / `translation_chunk` pairs (choose the minimum split that works)\n5. If splitting is risky, return `needs_manual_review`\n\n---\n\n## Absolute rules\n\n### 1) This is âsplittingâ, not âeditingâ\n\n- For the body text, use the input text as-is and, in principle, only take **substrings** from the provided `transliteration` / `translation`\n- Do not fill in missing text, paraphrase, summarize, or re-translate\n- Even if you notice OCR mistakes or unnatural English, **do not fix them in this task**\n\n### 2) Preserve the order\n\n- When you stack the chunks from top to bottom, they must preserve the original order of the `transliteration` / `translation`\n\n### 3) Only cut at boundaries visible on both sides\n\n- Sentence/clause boundaries on the English side\n- Boundaries between itemized lists\n- Breaks in letter templates / formulaic phrases\n- Breaks between imperative/conditional sections\n- Clear line-group boundaries on the transliteration side\n\nOnly cut at places where a coherent âunitâ is visible on both sides.\n\n### 4) Do not cut in the middle of these\n\n- Numbers, fractions, decimals, tax rates\n- Amount expressions like shekel / mina groupings\n- Personal names, divine names, place names\n- Phrases like `son of ...` or `belonging to ...`\n- Inside quotation marks\n- Inside an item name\n\n### 5) If itâs not safe, hold it\n\nIn the following cases, do not force a split; return `needs_manual_review`:\n\n- Even with the images, the aligned boundary cannot be determined\n- The English translation in train looks broken in the middle\n- The source/target are misaligned and 1:1 correspondence is unclear\n- Keeping it within 2â5 splits would inevitably break numeric/proper-noun sequences\n\n---\n\n## How to use the images\n\nUse the screenshots to decide **where to cut so that transliteration and translation stay aligned**.\n\nImportant:\n\n- The Akkadian and English text in the PDF may not exactly match the text provided in the chat.\n- This may be because the chat text has already been normalized (e.g., `<gap>` replacements, symbol cleanup, etc.).\n- Therefore, use the PDF as a reference to find alignment boundaries, but for the final output body text, **prefer the chat-provided text**.\n- Do not rewrite the body text to match the PDFâs visual appearance.\n\nPriority when using images:\n\n1. Verify alignment boundaries between `transliteration` and `translation`\n2. Resolve ambiguous boundaries that are hard to see from text alone\n3. Decide what spans form a coherent aligned unit\n\nDo not create new translations even if you use the images.\n\n---\n\n## Boundary heuristics (for a small number of splits)\n\n### Good candidates (high priority)\n\n- After an introductory formula like `a-na ... qĂ-bi-ma um-ma ...` (but donât split into a tiny chunk just for the intro)\n- English expressions that indicate a section change, e.g. `all this ...`, `thereof ...`, `therefore ...`, `if ... then ...`\n- Places where a large block of an item list changes (avoid splitting each small item)\n- Where repetition of the same pattern ends as a block\n\n### Tempting but usually forbidden\n\n- Splitting every English sentence (often creates too many chunks)\n- Splitting every small list item (avoid)\n\n---\n\n## Output format (strict)\n\nReturn **only a JSON array**. No extra explanation text.\n\n### When splitting is possible (2â5 chunks)\n\n```json\n[\n {\n \"record_id\": \"original record id\",\n \"split_id\": \"original record id__01\",\n \"status\": \"ok\",\n \"transliteration_chunk\": \"split transliteration chunk (must be a substring of the input)\",\n \"translation_chunk\": \"corresponding translation chunk (must be a substring of the input)\",\n \"boundary_reason\": \"one sentence describing why you cut here (alignment rationale)\",\n \"confidence\": \"high\"\n }\n]\n```\n\n### When splitting is risky\n\n```json\n[\n {\n \"record_id\": \"original record id\",\n \"split_id\": \"original record id__00\",\n \"status\": \"needs_manual_review\",\n \"transliteration_chunk\": \"\",\n \"translation_chunk\": \"\",\n \"boundary_reason\": \"cannot cut into 2â5 chunks without breaking alignment\",\n \"confidence\": \"low\"\n }\n]\n```\n\n---\n\n## `confidence` guide\n\n- `high`: boundaries are very clear from both images and text\n- `medium`: mostly reasonable, but alignment is slightly weak on one side\n- `low`: the original train text may be missing/broken/misaligned\n\n---\n\n## When you receive an actual request, always follow these\n\n- Return only a JSON array\n- Strongly prefer 2â5 splits and do not add too many boundaries\n- Remember this is splitting, not editing\n- Do not add words that are not in the provided body text, even if you look at images\n- If unsure, return `needs_manual_review`\nAt first I tried to clean up train.csv using Codex agents and the skills feature, but I couldnât control it well, and in the end I did it manually.
I started from the public notebook âLB 35.9 Ensembling & Post Processing Baselineâ, then increased the number of models and made small changes to preprocessing and postprocessing.\nThe last two submissions differed only in the number of training epochs; despite concerns about overfitting, the one trained for more epochs achieved a higher Private LB score.
\nBoth preprocessing and postprocessing are differences from âLB 35.9 Ensembling & Post Processing Baselineâ (only the differences are listed here).
\n_EXACT_FRAC_RE + _canon_decimal(float)):1/2 and 2 1/2 into Unicode vulgar fractions (e.g., ½)0.16666... â â
).5 values like 2.5 as-is (by design)Postprocessing and the final submission formatting are also differences from âLB 35.9 Ensembling & Post Processing Baselineâ (only the differences are listed here).
\n5/12 shekel(s): baseline â
shekel 15 grains â â
shekel 15 grains1/2 / 2 1/2 / decimals â Unicode fractions)you / she, keep the left side (you) and drop the right side (she)1/2\" / ') instead (baseline deletes them)() in bulk).. / xx..., etc.), and as a safeguard normalized ḍ/Ḫ to h/H<gap>validate_submission to check columns/row count/empty translations/duplicate ids (baseline mainly logs the number of empty outputs)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)train.csv; input length 1024, output length 1024)