| # Schema |
|
|
| One record per successful generation in `data/generations.jsonl`. |
|
|
| | Field | Type | Notes | |
| |---|---|---| |
| | `id` | int | Stable across releases. | |
| | `style` | string | Lettering style slug, e.g. `witch-font`. | |
| | `style_url` | string \| null | The style's page on aifont.ai; `null` for pilot and control styles that never went live. | |
| | `language` | string | `en` `de` `es` `pt` `zh` `zh-hant` `ja` `ar`. The prompt is written in this language. | |
| | `target_text` | string | The exact text the prompt asked the model to write. | |
| | `prompt_hash` | string | MD5 of the full prompt sent. Text in `data/prompts.jsonl`. | |
| | `model` | string | `GPT Image` · `Nano Banana` · `Grok Imagine` · `Seedream` · `Qwen-Image` | |
| | `model_id` | string | Upstream model version at generation time. | |
| | `resolution` / `quality` / `aspect_ratio` | string | Generation settings. | |
| | `generated_at` | string | UTC, ISO 8601. | |
| | `purpose` | string | `published` (shown on aifont.ai), `language_test`, `word_check`, `pilot`, `control`. | |
| | `attempt` | string | `first` · `retry` (an earlier image in the same cell was wrong) · `extra_sample` (earlier ones were right). Cell = style × language × model × prompt version. | |
| | `image_url` | string | Archived original, unedited. | |
| | `gallery_id` | int | The image's id on FontVibe, the service that generated it. | |
| | `text_verdict` | string | `correct` · `incorrect` · `unsure` | |
| | `error_types` | string[] | For `incorrect`: see below. | |
| | `rendered_text` | string \| null | What the image actually says, for spelling mistakes. | |
| | `effect_drift` | bool | The look missed the style badly enough that we regenerated. Independent of spelling. | |
| | `observations` | string[] | Recorded on correct images too; see below. | |
| | `reviews` | object[] | Each reviewer's latest call on the text: `reviewer` (`ai`, `maintainer`, or a contributor's GitHub handle), `text_verdict`, `native_speaker`, `zoomed`, `reviewed_at`. | |
| | `human_reviewed` | bool | At least one person reviewed this image. `false` means the AI's verdict only. | |
| | `disputed` | bool | Reviewers disagree and none is a native speaker. `text_verdict` is then `unsure`. | |
|
|
| ## `error_types` |
| |
| | Value | Meaning | |
| |---|---| |
| | `extra_letter` | An extra letter, stroke or character | |
| | `missing_letter` | A letter, stroke or character left out | |
| | `wrong_letter` | A different letter or character, often a look-alike | |
| | `order` | Letters out of order, including right-to-left scripts written backwards | |
| | `malformed` | Shapes that can't be read as any letter | |
| | `joining` | Wrong Arabic joining form (initial / medial / final) | |
| | `dots_diacritics` | Wrong dots or marks: Arabic dots, umlauts, accents, Japanese dakuten | |
| | `case` | Wrong case, only where the prompt asked for upper or lower case | |
| | `extra_text` | Text the prompt didn't ask for, competing with the target | |
| | `word_split` | The word broken apart or wrapped badly | |
|
|
| ## `observations` |
|
|
| `decoration_added` · `background_text` · `style_drift` · `layout_drift` · `color_drift` · `cropped` · `low_contrast` |
|
|
