Informational prompts t2i-10..14; retire t2i-04, t2i-09

prompts.jsonl

@@ -1,10 +1,13 @@
 {"id": "tgb-t2i-03", "modality": "text-to-image", "style": "crayon, hand-drawn banner", "prompt": "Create a wide banner image (roughly 3:1) drawn in crayon on textured paper, like a bright, playful hand-drawn explainer. At the top left, a box labelled \"Boxty v1.0\" with a small subtitle \"tech vocabulary\", and next to it, in handwriting: \"WER 1.35%\". At the top right, a box labelled \"Jachnun v1.0\" with a small subtitle \"code-switching\", and next to it, in handwriting: \"WER 5.85%\". Between them, a third box containing a jumble of words: \"teudat zehut\", \"arnona\", \"makolet\", \"GitHub Gist\", \"Kubernetes\". From that middle box, one bold arrow points straight down to a larger box at the bottom labelled \"Sabich v2.2\", with handwritten next to it: \"Code-switching WER 4.91%\" and \"Tech WER 2.76%\". In the margin beside the arrow, a scribbled note with a little doodle arrow pointing at Sabich: \"accuracy preserved!\". No other text.", "input_image": null, "expected_text": ["Boxty v1.0", "tech vocabulary", "WER 1.35%", "Jachnun v1.0", "code-switching", "WER 5.85%", "teudat zehut", "arnona", "makolet", "GitHub Gist", "Kubernetes", "Sabich v2.2", "Code-switching WER 4.91%", "Tech WER 2.76%", "accuracy preserved!"], "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "structure_checks": ["wide banner aspect ratio", "Boxty top left, Jachnun top right, word box between them", "each WER sits next to the right model", "exactly one arrow, from the word box down to Sabich v2.2", "margin note beside the arrow, pointing at Sabich", "crayon texture throughout", "no extra text"], "notes": "WER figures: Boxty v1.0 tech WER from its model card (hf.co/Hebrish/Parakeet-TDT-0.6B-v2-Tech-And-Accent, test revision f956dbb184c2); Jachnun v1.0 and Sabich v2.2 from the Sabich model card, English-Hebrew Code-Switching ASR Benchmark test set v2 (revision 48ec85a), checked 2026-10-28. Boxty's figure is on a different test set, so it is a label to reproduce, not a like-for-like comparison."}
-{"id": "tgb-t2i-04", "modality": "text-to-image", "style": "cover art, free style", "prompt": "Cover art for a blog post about fine-tuning a speech recognition model on one person's voice so it understands both tech jargon and Hebrew words dropped into English sentences. Two earlier specialist models (a tech-vocabulary model and a code-switching model) feed into one combined model that does both without getting worse at either. Show the idea visually: sound waves from a speaker turning into accurate text, two streams merging into one, a sense of things clicking into place. Warm, colourful and characterful, with a nod to Israeli street food (the models are named after foods: the combined one is Sabich, a stuffed pita). Leave clear space at the top for a title. Do not try to write a title or any sentences.", "input_image": null, "expected_text": [], "validation": [], "structure_checks": ["two streams visibly merge into one", "speech turning into text is legible as an idea", "clear space left at the top", "any text that appears is real, correctly spelled words; no pseudo-text or garbled glyphs"], "notes": "Descriptive prompt with no exact strings, no title and no diagram elements: ratings only, no validation ticks. structure_checks are a checklist for the ratings."}
 {"id": "tgb-t2i-05", "modality": "text-to-image", "style": "photoreal animal banner", "prompt": "Wide 16:9 banner image for a GitHub repository of home-server backup scripts. Photorealistic, warm natural light, shallow depth of field. A three-toed sloth sits at a cluttered desk in a small server closet, one long claw resting on a keyboard, looking calmly pleased with itself. The monitor in front of it shows a progress bar with the text \"Backup 3 of 7 complete\" above it. Behind the sloth, a rack of servers with blinking green lights, and a handwritten paper sign taped to the rack reading \"DO NOT UNPLUG\". On the desk, a white mug printed with \"NAS\". No other text, logos or lettering anywhere in the image.", "input_image": null, "expected_text": ["Backup 3 of 7 complete", "DO NOT UNPLUG", "NAS"], "validation": ["text_exact", "text_legible"], "structure_checks": ["photoreal sloth at a desk, claw on the keyboard", "progress bar on the monitor with the text above it", "paper sign taped to the server rack", "mug on the desk carries NAS", "no other text"], "notes": "Modelled on the author's sloth blog heroes (backupsloth, amazedslothgpts in Personal-Site-Blog)."}
 {"id": "tgb-t2i-06", "modality": "text-to-image", "style": "photoreal two-animal banner", "prompt": "Wide 16:9 banner image for a repository of command-line plugins for working with spreadsheets. Photorealistic, soft morning light from a kitchen window. A sloth and a capybara sit side by side at a wooden kitchen table, pair-programming on one open laptop. The sloth points at the screen with a long claw; the capybara looks sceptical. The laptop screen shows a dark terminal with a single line of large, clearly readable text: \"$ claude plugin install csv-tools\". Two coffee cups and a potted basil plant on the table. No other text, logos or lettering anywhere in the image.", "input_image": null, "expected_text": ["$ claude plugin install csv-tools"], "validation": ["text_exact", "text_legible"], "structure_checks": ["two photoreal animals: a sloth and a capybara", "sloth points at the screen, capybara looks sceptical", "terminal line readable on the laptop screen", "no other text"], "notes": "One exact string, but small and on a screen inside a photoreal scene: the hardest place for text in a banner."}
 {"id": "tgb-t2i-07", "modality": "text-to-image", "style": "mascot in a photoreal scene", "prompt": "Wide 16:9 banner image for a plugin that looks up bus arrival times. Photorealistic street photograph, late-afternoon sun, Jerusalem stone buildings. A small, cute, blocky terracotta-orange robot mascot (soft felt texture, square body, two dot eyes, stubby arms) stands at a bus stop, looking up at an electronic arrival board. The board is amber LED on black and shows exactly two rows, in this order: first row \"18  Central Station  4 min\", second row \"74  Malha Mall  11 min\". A city bus is pulling in behind the mascot. No other text, logos or lettering anywhere in the image.", "input_image": null, "expected_text": ["18", "Central Station", "4 min", "74", "Malha Mall", "11 min"], "validation": ["text_exact", "text_legible"], "structure_checks": ["mascot is a stylised character placed in a photoreal scene", "arrival board has exactly two rows", "row order: 18 first, 74 second, each with its own destination and time", "bus behind the mascot", "no other text (including on the bus)"], "notes": "Modelled on the author's Claude plugin banners (blocky mascot in a real Israeli street or airport). A real Jerusalem bus would carry Hebrew; the prompt forbids it so any lettering on the bus counts as invented text."}
 {"id": "tgb-t2i-08", "modality": "text-to-image", "style": "neon title banner", "prompt": "Wide 16:9 title banner for a GitHub README. A glowing neon sign mounted on a dark wall covered in tangled coloured cables and small circuit boards, retro arcade feel, with small pixel-art accents in the corners. The sign reads, on two lines, in a rounded bold typeface: first line \"Daniel's MCP\" glowing green-to-cyan, second line \"server collection\" glowing orange-to-pink. A small pixel-art sloth face sits centred above the text. No other text, logos or lettering anywhere in the image.", "input_image": null, "expected_text": ["Daniel's MCP", "server collection"], "validation": ["text_exact", "text_legible"], "structure_checks": ["two lines in the stated order", "line colours as asked (green-to-cyan, then orange-to-pink)", "pixel-art sloth face centred above", "no other text, including on the circuit boards"], "notes": "Modelled on the Claude-Code-Plugins README banner. The apostrophe counts toward text_exact."}
-{"id": "tgb-t2i-09", "modality": "text-to-image", "style": "photoreal animal hero, descriptive", "prompt": "Hero image for a blog post about running large language models on a GPU at home instead of paying for cloud APIs. Photorealistic and a little funny: a sloth in a cosy home office at night, lit by the glow of a large desktop PC with a big graphics card visible through its glass side panel, fans spinning. The sloth looks proud and slightly overheated. Through the window, a faint city skyline. Leave clear, uncluttered space on the left third for a title to be added later. Do not write any text, logos, labels or numbers anywhere in the image.", "input_image": null, "expected_text": [], "validation": [], "structure_checks": ["photoreal sloth with a home GPU rig", "left third left clear for a title", "any text that appears fails: no logos, brand names or pseudo-text on the PC or screen"], "notes": "Descriptive prompt with no exact strings, no title and no diagram elements: ratings only, no validation ticks. Invented text or brand names on the PC count against the ratings."}
+{"id": "tgb-t2i-10", "modality": "text-to-image", "style": "editorial lineage diagram", "prompt": "A 16:9 model-lineage diagram for a blog post. Calm editorial style for a personal tech blog: warm cream background, generous white space, thin slate-grey lines, a muted palette of deep teal, amber and slate, clean modern sans-serif text with a serif for the heading. Flat, no 3D, no gradients. Serif heading at the top left: \"One base, two recipes\". On the left, one box labelled \"Whisper Small\" with a smaller grey line under it: \"stock model\". From that box two arrows run to the right, one up to an upper row and one down to a lower row. Upper row, three small pale boxes joined by plus signs, in this order: \"Tech terms\", \"Own voice\", \"Replay\", then an arrow to a teal badge reading \"Boxty v1.0\" with \"210 min\" in small text under it. Lower row, in this order: \"Code-switching\", \"Own voice\", \"Replay\", then an arrow to an amber badge reading \"Jachnun v1.0\" with \"483 min\" under it. No other text, logos or watermarks anywhere.", "input_image": null, "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "expected_text": ["One base, two recipes", "Whisper Small", "stock model", "Tech terms", "Own voice", "Replay", "Boxty v1.0", "210 min", "Code-switching", "Jachnun v1.0", "483 min"], "structure_checks": ["one base box on the left", "two arrows from it: one to the upper row, one to the lower row", "upper row order: Tech terms + Own voice + Replay -> Boxty v1.0", "lower row order: Code-switching + Own voice + Replay -> Jachnun v1.0", "Boxty badge teal, Jachnun badge amber, minutes under the right badge"], "notes": "Modelled on the lineage figure in the Jachnun/Sabich post on blog.danielrosehill.com. 'Own voice' and 'Replay' appear twice each; text_exact needs both copies. Minutes are illustrative."}
+{"id": "tgb-t2i-11", "modality": "text-to-image", "style": "editorial title card", "prompt": "A 16:9 social share card for a blog post. Calm editorial style for a personal tech blog: warm cream background, generous white space, thin slate-grey lines, a muted palette of deep teal, amber and slate, clean modern sans-serif text with a serif for the heading. Flat, no 3D, no gradients. Large serif title at the top left: \"Sabich v2.2\". Under it, in smaller sans-serif: \"One model for tech jargon and Hebrew words.\" Below that, a white rounded card with a thin border holding two rows. Row one: a small grey label \"STOCK\" on the left, and in muted red: \"I need to renew my to the zehut\". Row two: a small teal label \"FINE-TUNED\" on the left, and in bold teal: \"I need to renew my teudat zehut\". No other text, logos or watermarks anywhere.", "input_image": null, "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "expected_text": ["Sabich v2.2", "One model for tech jargon and Hebrew words.", "STOCK", "I need to renew my to the zehut", "FINE-TUNED", "I need to renew my teudat zehut"], "structure_checks": ["title, subtitle, then the card, top to bottom", "STOCK row above FINE-TUNED row", "stock line in red, fine-tuned line bold teal", "the deliberately wrong 'to the zehut' kept as written, not corrected"], "notes": "Modelled on the Hebrish share card (STOCK vs HEBRISH rows) on blog.danielrosehill.com. The stock line is a deliberate mis-transcription: a model that 'fixes' it fails text_exact."}
+{"id": "tgb-t2i-12", "modality": "text-to-image", "style": "information table", "prompt": "A 16:9 information table drawn as a clean graphic for a blog post. Calm editorial style for a personal tech blog: warm cream background, generous white space, thin slate-grey lines, a muted palette of deep teal, amber and slate, clean modern sans-serif text with a serif for the heading. Flat, no 3D, no gradients. Serif title above the table: \"Hebrish models at a glance\". Four columns with a deep-teal header row in white text, in this order: \"Model\", \"Trained on\", \"Tech WER\", \"Code-switching WER\". Three body rows with thin rules between them, in this order: row 1 \"Boxty v1.0\", \"Tech vocabulary\", \"1.35%\", \"9.80%\"; row 2 \"Jachnun v1.0\", \"Code-switching\", \"3.12%\", \"5.85%\"; row 3 \"Sabich v2.2\", \"Both\", \"2.76%\", \"4.91%\". Highlight the Sabich v2.2 row with a pale amber fill. Numbers right-aligned. Under the table, in small grey text: \"Lower is better.\" No other text, logos or watermarks anywhere.", "input_image": null, "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "expected_text": ["Hebrish models at a glance", "Model", "Trained on", "Tech WER", "Code-switching WER", "Boxty v1.0", "Tech vocabulary", "1.35%", "9.80%", "Jachnun v1.0", "Code-switching", "3.12%", "5.85%", "Sabich v2.2", "Both", "2.76%", "4.91%", "Lower is better."], "structure_checks": ["exactly 4 columns in the stated order", "exactly 3 body rows in the stated order", "every value in its own row and column (no swapped cells)", "teal header, amber Sabich row", "no extra rows, columns or totals"], "notes": "Tables are a common blog ask. 1.35, 5.85, 2.76 and 4.91 are the real figures from tgb-t2i-03; 9.80 and 3.12 are illustrative. A cell in the wrong column fails structure_ok even if spelled right."}
+{"id": "tgb-t2i-13", "modality": "text-to-image", "style": "numbered-steps infographic", "prompt": "A 16:9 infographic for a blog post showing a five-step process from left to right. Calm editorial style for a personal tech blog: warm cream background, generous white space, thin slate-grey lines, a muted palette of deep teal, amber and slate, clean modern sans-serif text with a serif for the heading. Flat, no 3D, no gradients. Serif title at the top: \"From voice note to blog post\". Five numbered circles in a row, numbered 1 to 5, joined by thin arrows pointing right. Above each circle a simple flat line icon; below each, a bold label and a one-line caption in grey. In order: 1 \"Record\" / \"Phone voice memo\"; 2 \"Transcribe\" / \"Fine-tuned ASR\"; 3 \"Clean up\" / \"Paragraphs and punctuation\"; 4 \"Draft\" / \"LLM with my style guide\"; 5 \"Publish\" / \"Astro site on Vercel\". No other text, logos or watermarks anywhere.", "input_image": null, "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "expected_text": ["From voice note to blog post", "1", "2", "3", "4", "5", "Record", "Phone voice memo", "Transcribe", "Fine-tuned ASR", "Clean up", "Paragraphs and punctuation", "Draft", "LLM with my style guide", "Publish", "Astro site on Vercel"], "structure_checks": ["five steps, numbered 1-5 left to right", "arrows between consecutive steps, all pointing right", "each caption under its own label", "an icon above each circle that fits the step"], "notes": "Numbered-step infographics are the other common informational ask. Icons with embedded text (e.g. a fake app name) count as invented text."}
+{"id": "tgb-t2i-14", "modality": "text-to-image", "style": "bar chart", "prompt": "A 16:9 horizontal bar chart for a blog post. Calm editorial style for a personal tech blog: warm cream background, generous white space, thin slate-grey lines, a muted palette of deep teal, amber and slate, clean modern sans-serif text with a serif for the heading. Flat, no 3D, no gradients. Serif title at the top left: \"Word error rate on tech vocabulary\". Five horizontal bars, top to bottom in this order, each with its label on the left and its value at the end of the bar: \"Whisper Small (stock)\" 8.4%, \"Whisper Medium (stock)\" 6.1%, \"Parakeet TDT (stock)\" 4.9%, \"Sabich v2.2\" 2.76%, \"Boxty v1.0\" 1.35%. Bar lengths proportional to the values on a shared axis starting at zero. Stock models in slate grey, the two fine-tunes in teal. Light vertical gridlines; x-axis label under the chart: \"WER (%)\". No other text, logos or watermarks anywhere.", "input_image": null, "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "expected_text": ["Word error rate on tech vocabulary", "Whisper Small (stock)", "8.4%", "Whisper Medium (stock)", "6.1%", "Parakeet TDT (stock)", "4.9%", "Sabich v2.2", "2.76%", "Boxty v1.0", "1.35%", "WER (%)"], "structure_checks": ["five bars in the stated order, longest at the top", "bar lengths roughly proportional (8.4 about six times 1.35)", "each value at the end of its own bar", "stock bars grey, fine-tune bars teal", "axis starts at zero; any tick numbers must be correct"], "notes": "Charts from an image model are a stretch, but the blog uses them (Knesset forecast dot plot). Tick numbers are allowed if correct; wrong ticks fail text_exact. Stock figures are illustrative."}
 {"id": "tgb-i2i-01", "modality": "image-to-image", "style": "keep handwriting, add colour", "prompt": "Turn this whiteboard photo into a clean, colourful diagram. Remove the whiteboard frame, glare and background, and correct the perspective to a straight-on view on a white background. Keep my handwriting style but make it neater and easy to read. Keep every label, box and arrow exactly where it is and pointing the same way. Fix obvious misspellings of technical names (\"Proxmox\", \"Replicate\", \"Remote\") but do not add, remove or reword any label. Add colour: give each group of related boxes its own soft colour and make the arrows a contrasting colour.", "input_image": "inputs/whiteboard-mcp-01.jpg", "expected_text": ["WAN", "Telegram", "CF / Tailscale", "OpenClaw", "Docker", "VM", "META MCP", "META MCP 2", "Desktop Client", "No proxy possible", "LAN use", "LAN MCPs", "NAS", "HA", "SBCs", "Proxmox", "SSE Remote MCPs", "Replicate", "Google"], "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "structure_checks": ["Desktop Client -> META MCP 2", "Desktop Client -> VM", "WAN -> OpenClaw (labelled CF / Tailscale)", "META MCP -> LAN MCPs", "META MCP -> SSE Remote MCPs", "LAN MCPs -> NAS, HA, SBCs, Proxmox", "SSE Remote MCPs -> Replicate, Google", "upward arrow labelled LAN use", "handwriting character kept", "no labels added or dropped"]}
 {"id": "tgb-i2i-03", "modality": "image-to-image", "style": "chalkboard", "prompt": "Turn this whiteboard photo into a diagram drawn on a classic green chalkboard. Remove the physical whiteboard, its frame, the pen tray, glare and the wall. Correct the perspective so the result is a straight-on view. Use white chalk for the text and coloured chalk (yellow, pink, light blue) for the boxes and arrows, with subtle chalk dust and smudges. The lettering should be neat chalk handwriting. Preserve every word, box and arrow, and do not add, remove or reword any text: copy each word as written, correcting only what the handwriting makes ambiguous. Keep every arrow pointing the same way between the same items.", "input_image": "inputs/whiteboards-08.jpg", "expected_text": ["Few Shot Learning - 3", "Zoom Meeting Prep SOP", "Is my hair a mess?", "Y", "N", "COMB HAIR", "CHILL OUT!"], "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "structure_checks": ["question branches into Y on the left and N on the right", "Y -> COMB HAIR, N -> CHILL OUT!", "COMB HAIR and CHILL OUT! are boxed", "title underlined at top left", "green chalkboard look"], "notes": "Input: https://huggingface.co/datasets/danielrosehill/Whiteboards sample 08 (flowchart). expected_text from its human transcription; case-insensitive (the board mixes caps)."}
 {"id": "tgb-i2i-04", "modality": "image-to-image", "style": "sticky notes", "prompt": "Turn this whiteboard photo into a colourful sticky-note board. Remove the physical whiteboard, its frame, the pen tray, glare and the wall. Correct the perspective so the result is a straight-on view. Put it on a light cork-board background. Render each block of text and each box as its own sticky note (yellow, pink, blue, green) with a slight shadow and tilt, in a casual handwritten font, and draw the arrows as marker lines between the notes. Keep the little bot and smiley faces as small doodles. Preserve every word, box and arrow, and do not add, remove or reword any text: copy each word as written, correcting only what the handwriting makes ambiguous. Keep every arrow pointing the same way between the same items.", "input_image": "inputs/whiteboards-11.jpg", "expected_text": ["PERSONALISED AI", "BASIC IDEA", "Use-case:", "Human wants to create a task-specific AI agent", "& frontload context", "Tactic:", "Create AI agent to \"interview\" you", "transcribe", "(could be voice → voice!)", "extract content", "vector DB", "What's your fav food?", "PIZZA!", "CREATE MEMORY", "Daniel's fav food = PIZZA", "VECTOR DB"], "validation": ["text_exact", "text_legible", "structure_ok", "meaning_preserved"], "structure_checks": ["bot face -> speech bubble What's your fav food?", "PIZZA! reply next to the smiley face", "bot -> CREATE MEMORY box, PIZZA! -> memory box", "memory box -> VECTOR DB cylinder", "tactic line keeps its arrow chain in order", "bot and smiley faces kept"], "notes": "Input: https://huggingface.co/datasets/danielrosehill/Whiteboards sample 11 (diagram). 16 strings; case-insensitive. A dense board: tests whether splitting text across notes drops or rewords any of it."}

retired.jsonl

@@ -1,3 +1,5 @@
 {"id": "tgb-t2i-01", "modality": "text-to-image", "style": "anime / manga", "prompt": "Create a bright, colourful anime-style explainer diagram for a blog post about fine-tuning speech recognition models for domain vocabulary. Title at the top, exactly: \"Fine-Tuning for Domain Vocabulary\". Two side-by-side panels. Left panel heading: \"Whisper Small\". Right panel heading: \"Whisper Large\". In each panel, a box labelled \"Base Model\" on the left with an arrow pointing right to a box labelled \"Fine-Tuned Model\". Above each arrow, the label \"+ 2 hrs domain audio\". Under the left panel's boxes: \"WER 18.4%\" under Base Model and \"WER 7.1%\" under Fine-Tuned Model. Under the right panel's boxes: \"WER 9.6%\" under Base Model and \"WER 5.2%\" under Fine-Tuned Model. At the bottom, a caption banner reading exactly: \"Smaller models gain more from fine-tuning\". A cheerful anime character in the corner points at the left panel. No other text.", "input_image": null, "expected_text": ["Fine-Tuning for Domain Vocabulary", "Whisper Small", "Whisper Large", "Base Model", "Fine-Tuned Model", "+ 2 hrs domain audio", "WER 18.4%", "WER 7.1%", "WER 9.6%", "WER 5.2%", "Smaller models gain more from fine-tuning"], "structure_checks": ["two side-by-side panels", "in each panel an arrow from Base Model to Fine-Tuned Model", "WER figures sit under the correct box in the correct panel", "character points at the left panel", "no extra text"], "retired": "Retired 2026-10-28: the anime/manga style ask doesn't match how the benchmark author actually illustrates repos and posts (photoreal banners, often with sloths or other animals; a blocky mascot in a real scene; neon or crayon title art). Both saved annotations of round 2026-10-28-informal were on these prompts and scored rating_overall 2/10 despite passing every validation tick: 'looks like something you'd hang up in a kindergarten instead of put on a tech blog' (tgb-t2i-01, Ideogram V4.5); 'functional diagram, just not terribly exciting' (tgb-t2i-02, FLUX.3 Klein). Those annotations stay in the ratings dataset."}
 {"id": "tgb-t2i-02", "modality": "text-to-image", "style": "manga, bold ink with flat colour", "prompt": "Create a manga-style technical diagram, bold ink lines with flat bright colours, showing a speech-to-text pipeline. Four boxes in a single left-to-right row, connected by arrows pointing right: \"Audio In\" -> \"Voice Activity Detection\" -> \"Transcription\" -> \"Text Out\". Below the row, a separate box labelled \"Custom Vocabulary\" with one arrow pointing up into \"Transcription\" only. A small speech bubble coming out of \"Text Out\" says exactly: \"Kubernetes, not Cooper Nettys!\". No other text.", "input_image": null, "expected_text": ["Audio In", "Voice Activity Detection", "Transcription", "Text Out", "Custom Vocabulary", "Kubernetes, not Cooper Nettys!"], "structure_checks": ["four boxes in one row, left to right, in the stated order", "three right-pointing arrows between them", "Custom Vocabulary sits below the row", "exactly one arrow from Custom Vocabulary, pointing into Transcription", "speech bubble comes from Text Out", "no extra text"], "retired": "Retired 2026-10-28: the anime/manga style ask doesn't match how the benchmark author actually illustrates repos and posts (photoreal banners, often with sloths or other animals; a blocky mascot in a real scene; neon or crayon title art). Both saved annotations of round 2026-10-28-informal were on these prompts and scored rating_overall 2/10 despite passing every validation tick: 'looks like something you'd hang up in a kindergarten instead of put on a tech blog' (tgb-t2i-01, Ideogram V4.5); 'functional diagram, just not terribly exciting' (tgb-t2i-02, FLUX.3 Klein). Those annotations stay in the ratings dataset."}
 {"id": "tgb-i2i-02", "modality": "image-to-image", "style": "anime, clean sans-serif labels", "prompt": "Redraw this whiteboard sketch as a vibrant anime-style infographic. Replace all of my handwritten labels with a clean sans-serif font, spelled correctly. Keep the same layout: every label, box and arrow stays in the same relative position and every arrow points the same way. Draw the two little faces as small cute anime characters, keeping their labels (\"Desktop Client\" and \"Telegram\"). Do not add, remove or reword any label.", "input_image": "inputs/whiteboard-mcp-01.jpg", "expected_text": ["WAN", "Telegram", "CF / Tailscale", "OpenClaw", "Docker", "VM", "META MCP", "META MCP 2", "Desktop Client", "No proxy possible", "LAN use", "LAN MCPs", "NAS", "HA", "SBCs", "Proxmox", "SSE Remote MCPs", "Replicate", "Google"], "structure_checks": ["Desktop Client -> META MCP 2", "Desktop Client -> VM", "WAN -> OpenClaw (labelled CF / Tailscale)", "META MCP -> LAN MCPs", "META MCP -> SSE Remote MCPs", "LAN MCPs -> NAS, HA, SBCs, Proxmox", "SSE Remote MCPs -> Replicate, Google", "upward arrow labelled LAN use", "all labels in sans-serif type", "no labels added or dropped"], "retired": "Retired 2026-10-28: the anime/manga style ask doesn't match how the benchmark author actually illustrates repos and posts (photoreal banners, often with sloths or other animals; a blocky mascot in a real scene; neon or crayon title art). Both saved annotations of round 2026-10-28-informal were on these prompts and scored rating_overall 2/10 despite passing every validation tick: 'looks like something you'd hang up in a kindergarten instead of put on a tech blog' (tgb-t2i-01, Ideogram V4.5); 'functional diagram, just not terribly exciting' (tgb-t2i-02, FLUX.3 Klein). Those annotations stay in the ratings dataset."}
+{"id": "tgb-t2i-04", "modality": "text-to-image", "style": "cover art, free style", "prompt": "Cover art for a blog post about fine-tuning a speech recognition model on one person's voice so it understands both tech jargon and Hebrew words dropped into English sentences. Two earlier specialist models (a tech-vocabulary model and a code-switching model) feed into one combined model that does both without getting worse at either. Show the idea visually: sound waves from a speaker turning into accurate text, two streams merging into one, a sense of things clicking into place. Warm, colourful and characterful, with a nod to Israeli street food (the models are named after foods: the combined one is Sabich, a stuffed pita). Leave clear space at the top for a title. Do not try to write a title or any sentences.", "input_image": null, "expected_text": [], "validation": [], "structure_checks": ["two streams visibly merge into one", "speech turning into text is legible as an idea", "clear space left at the top", "any text that appears is real, correctly spelled words; no pseudo-text or garbled glyphs"], "notes": "Descriptive prompt with no exact strings, no title and no diagram elements: ratings only, no validation ticks. structure_checks are a checklist for the ratings.", "retired": "Retired 2026-10-28: a descriptive prompt with no text, title or diagram asked for gives nothing to validate, so it doesn't belong in a text-and-structure benchmark (annotator, 2026-10-28: 'doesn't meet the validation criteria for the benchmark itself'). Annotations already made stay in the ratings dataset."}
+{"id": "tgb-t2i-09", "modality": "text-to-image", "style": "photoreal animal hero, descriptive", "prompt": "Hero image for a blog post about running large language models on a GPU at home instead of paying for cloud APIs. Photorealistic and a little funny: a sloth in a cosy home office at night, lit by the glow of a large desktop PC with a big graphics card visible through its glass side panel, fans spinning. The sloth looks proud and slightly overheated. Through the window, a faint city skyline. Leave clear, uncluttered space on the left third for a title to be added later. Do not write any text, logos, labels or numbers anywhere in the image.", "input_image": null, "expected_text": [], "validation": [], "structure_checks": ["photoreal sloth with a home GPU rig", "left third left clear for a title", "any text that appears fails: no logos, brand names or pseudo-text on the PC or screen"], "notes": "Descriptive prompt with no exact strings, no title and no diagram elements: ratings only, no validation ticks. Invented text or brand names on the PC count against the ratings.", "retired": "Retired 2026-10-28: a descriptive prompt with no text, title or diagram asked for gives nothing to validate, so it doesn't belong in a text-and-structure benchmark (annotator, 2026-10-28: 'doesn't meet the validation criteria for the benchmark itself'). Annotations already made stay in the ratings dataset."}