| --- |
| license: other |
| task_categories: |
| - visual-question-answering |
| language: |
| - en |
| --- |
| |
| # rects_train_cleaned |
|
|
| The **rects_train** family of the ElliotVL supervised-fine-tuning pool, **after VLM cleaning**. |
| |
| | | | |
| |---|---| |
| | images | 40,563 | |
| | QA turns | 282,249 | |
| | answers rewritten by the cleaning pass | 0 | |
| | QA created by the cleaning pass (`new_qa`) | not measured for this family | |
| | shards | 32 | |
| |
| ## How this was cleaned |
| |
| A vision-language model read each image together with its QA and judged the item. The pass is |
| **not a filter that only removes rows** — it rewrites answers it finds wrong but salvageable, |
| drops what it cannot salvage, and adds QA where the image supports more than the source provided. |
| Each row carries the judge's own record in `clean_meta`, including the cleaner identity, the |
| policy it applied, and its per-item scores for legibility, richness and coverage. |
|
|
| **A large share of the QA here was written by the cleaning pass, not by the original dataset.** |
| Across the pool that share runs from roughly half to over 80% of a family's turns, and it is |
| reported in the table above. Those rows have an empty `org_answer` because no pre-clean original |
| exists. Treat them as model-generated supervision: they were judged against the image, but they |
| are not human annotation, and model-written QA is where formatting defects are most likely. |
|
|
| The effect on the answers that *were* carried over is substantive rather than cosmetic. In this pool the pass corrects values, not just |
| wording — a curve's critical point restated from `4.00` to `2.00`, a computed ratio from `1` to |
| `Approximately 1.33` — and for some families it removed the content entirely, which is why those |
| families are absent here rather than published empty. |
|
|
| ## `answer` vs `org_answer` |
| |
| - `answer` — the cleaned answer. **Train on this.** |
| - `org_answer` — the pre-cleaning answer from the same `(image, question)` in the uncleaned pool. |
| It is empty for QA the cleaning pass added, which have no original. |
|
|
| > For this family `org_answer` is empty throughout: recovering it means scanning the whole pre-clean family, which is raw family is 17 GB (> 6 GB cap). The cleaned `answer` is unaffected. |
| |
| Where the two differ, the difference is the correction. Keeping both makes every edit auditable |
| instead of asking you to take the cleaning on trust. |
| |
| |
| |
| ## Schema |
| |
| - `image` — HF `Image()`, renders directly in the dataset viewer |
| - `image_sha256` — content hash of the image |
| - `n_qa` — number of turns attached to this image |
| - `qa` — list of {`question`, `answer`, `org_answer`}, **in source order** |
| - `family` / `clean_meta` — provenance and the judge's per-row record |
|
|
| A row with more than one entry in `qa` is a **conversation over a single image**, not a set of |
| independent questions — the turns depend on each other and the image belongs to all of them. |
| Keep them together and in order when training. |
|
|
| Parquet row groups are written at ~32 MB so the dataset viewer pages without stalling. |
|
|
| ## What is NOT claimed |
|
|
| No rejection sampling was run against this family: there is no accept/reject verdict per row, and |
| `answer` is the cleaned reference rather than a model generation that a string verifier confirmed. |
| For that, use the `*_RS_think` / `*_rejected` datasets in [elliot-mllm](https://huggingface.co/elliot-mllm). |
|
|
|
|
| ## Examples |
| 10 rows taken straight from this dataset. Images are the **original files** as stored here - no downscaling, no recompression - and every turn is shown in full. |
|
|
| ### Example 1 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Please extract all text from the image. | 臻选 哆吉客 现制汉堡更新鲜！ DIGIKe | |
| | 2 | Describe the color scheme and layout of the main storefront sign. | The sign features a large orange rectangular panel on top containing the logo and main Chinese text, sitting above a narrower black strip containing the slogan and English brand name. | |
| | 3 | What does the logo on the left side of the orange sign look like? | It is a white, stylized circular graphic that resembles a simplified burger or a face with a dot inside. | |
| | 4 | What lighting fixtures are visible beneath the black strip of the sign? | There are three small, round recessed lights visible in the dark area below the sign. | |
|
|
|
|
| ### Example 2 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Could you extract all the text contained in the picture? | 品长 ® CHINESE FAST FOOD 中式快餐 加盟热线:133-13888813 送餐电话:82026918 | |
| | 2 | What are the visual characteristics of the main signboard background? | The signboard has a red background composed of horizontal slats. | |
| | 3 | What architectural features are visible directly above the red sign? | There is a brick wall with windows; the central window is covered with a translucent plastic sheet. | |
| | 4 | What is located directly beneath the red signboard? | A brown roll-down metal shutter. | |
|
|
|
|
| ### Example 3 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Please retrieve all the textual content from the figure. | 肚子里有料 香港爆汁流沙包 | |
| | 2 | What is the background color of the illuminated sign? | Yellow | |
| | 3 | What color is the smaller text at the bottom of the sign? | Red | |
| | 4 | What is visible on the far right edge of the image? | A green light | |
|
|
|
|
| ### Example 4 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Kindly pull out all the text present in the figure. | 馬記 大锅煮 奥城店 | |
| | 2 | What is the background color of the vertical text '馬記'? | Red. | |
| | 3 | What is the background material of the sign? | Horizontal wooden slats. | |
| | 4 | What graphic element is stylized into the middle character '锅'? | Chopsticks or skewers. | |
|
|
|
|
| ### Example 5 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Could you extract every piece of text from the picture? | 趣果时光 欢迎光临 试营业 鲜果水吧 时光 享·健康 趣果时光 | |
| | 2 | What material covers the lower part of the storefront wall and extends onto the floor? | Green artificial turf | |
| | 3 | What object is posted in the center of the window above the green turf? | A framed menu or poster | |
| | 4 | What vertical object stands on the far right side of the image? | A white pole | |
|
|
|
|
| ### Example 6 — 3 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Could you extract all the text contained in the picture? | 8 顾家凉皮 Since2008 入口绵 吃后想 GUJIA 顾 凉 皮 烧烤 肉夹馍 60 | |
| | 2 | What graphical icons are displayed next to the menu items on the far right of the sign? | There are three circular icons: a bowl (next to 凉皮), skewers (next to 烧烤), and a bun (next to 肉夹馍). | |
| | 3 | What is visible below the main illuminated sign? | There are menu boards displaying food items and prices, including a visible price of '60'. | |
|
|
|
|
| ### Example 7 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Please extract all text from the image. | 心情故事 Mood story 咖啡厅 咖啡 茶饮 休闲 | |
| | 2 | What graphic icon is visible on the left side of the sign? | A steaming coffee cup. | |
| | 3 | What is the background color of the main signboard? | Blue. | |
| | 4 | What is visible at the bottom of the image below the main sign? | A doorway or window looking into the shop, illuminated by two bright lights. | |
|
|
|
|
| ### Example 8 — 5 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Kindly extract all the written information from the image. | 小 凤爪王 嘉 | |
| | 2 | What is the background texture and color of the main signboard? | Dark brown horizontal slats. | |
| | 3 | How is the signboard illuminated? | Rows of small light bulbs line the top and bottom edges of the sign. | |
| | 4 | Describe the logo located to the left of the main text. | A red circle containing a yellow symbol. | |
| | 5 | What are the colors of the characters on the sign? | The main characters '凤爪王' are white, while the characters '小' and '嘉' at the bottom are red. | |
|
|
|
|
| ### Example 9 — 5 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Kindly pull out all the text present in the figure. | 福禧得农贸市场 宜家乐生鲜 晨缘轩包子店 水果店 | |
| | 2 | What is the text on the top line of the green sign? | 福禧得农贸市场 | |
| | 3 | What is the text on the middle line of the green sign? | 宜家乐生鲜 | |
| | 4 | What is the text on the bottom line of the green sign? | 晨缘轩包子店 | |
| | 5 | What text is visible on the glass inside the shop on the far right? | 水果店 | |
|
|
|
|
| ### Example 10 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | What treatment is being referred to in this context? | Rising Damp Treatment. | |
| | 2 | How is the text 'RISING DAMP' visually styled differently from the word 'TREATMENT'? | 'RISING DAMP' is white text inside a solid red rectangular box, whereas 'TREATMENT' is white text placed directly on the wall background without a box. | |
| | 3 | Describe the condition of the wall surface shown in the background. | The wall shows significant damage with peeling paint and plaster, revealing dark, damp-looking patches and brownish underlying material, particularly concentrated near the bottom. | |
| | 4 | What architectural feature is visible along the very bottom edge of the image? | A row of red bricks is visible at the bottom. | |
|
|
|
|
