| --- |
| license: other |
| task_categories: |
| - visual-question-answering |
| language: |
| - en |
| --- |
| |
| # synthdog_cleaned |
| |
| The **synthdog__x** family of the ElliotVL supervised-fine-tuning pool, **after VLM cleaning**. |
| |
| | | | |
| |---|---| |
| | images | 445,694 | |
| | QA turns | 1,613,204 | |
| | answers rewritten by the cleaning pass | 0 | |
| | QA created by the cleaning pass (`new_qa`) | not measured for this family | |
| | shards | 547 | |
|
|
| ## 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 312 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 identify and transcribe the text found in this image. | Dares Wins Vol. 5 Tommy's Heroes Vol . 6: For Tomorrow Vol. 7: Closin g Time miniserie s.Clark Kent is being inter viewed about Su perm an 's connectio n to notorious kille r Tommy Monag han. Taki ng the co nversation o | |
| | 2 | What color is the word 'Heroes' printed on the paper? | Blue | |
| | 3 | What is the paper resting on? | Green grass | |
| | 4 | What blue object is visible in the background on the right? | A blue plastic crate or bin | |
|
|
|
|
| ### Example 2 — 3 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | For documentation purposes, I need the text within this image transcribed. Could you provide that? | ccer, softball, and volleyball.The Lad y Buccaneers and Buccane ers have won multiple athletic competitions i ncluding the 2002 Divisio n II women's soccer cham pionship and the 2008 G SC me n's bask etball cha mpi onship.The Men's Soccer Team has won back to back confere nce titles under coach Clint Browne during the 2011 | |
| | 2 | How is the text spatially arranged on the page? | The text is fragmented into distinct sections: a main block in the upper center, a vertical stack of words on the right side ('SC me n's bask etball cha mpi'), a line below the main block starting with 'onship', and two lines at the bottom ('nce titles...' and 'Browne...'). | |
| | 3 | What is the background surface visible behind the paper? | A textured, light-colored surface that appears to be fabric. | |
|
|
|
|
| ### Example 3 — 5 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Please read and relay the text visible in this image. What does it say? | m to North Korea as part of the Agreed Framework. In releasing the funding President George W Bush waived the Framework's requirement that North Korea allow inspecto | |
| | 2 | How is the text arranged on the page? | It is arranged in two columns. | |
| | 3 | What is the style of the font used in the text? | The text is in an italic serif font. | |
| | 4 | Are there any words printed in colors other than black? | Yes, 'part of t' is reddish-brown and 'George' is purple. | |
| | 5 | Is the text in the image a complete sentence? | No, it appears to be a fragment starting with 'm' and ending with 'inspecto'. | |
|
|
|
|
| ### Example 4 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Please identify and transcribe the text found in this image. | osure exp eriment. his time outsi de tot aled 3 h ours and 8 minut es, bu t th e tr oub les were not all ov er: Af ter clo sin g the Kvant 2 EVA hatch , the cosmonauts f ound a 2 m m gap i | |
| | 2 | What is visible in the background behind the white board? | A person holding a white dog, a red chair, a house with siding, trees, and grass. | |
| | 3 | Which words in the text are colored purple or blue? | "t th" and "Af" are purple. "e tr" is blue. "minut" is dark blue. | |
| | 4 | What font style is used for the smaller text at the bottom of the board? | Italic. | |
|
|
|
|
| ### Example 5 — 3 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | There's text embedded in this image that I need extracted. Can you transcribe it verbatim? | ar. In th e genera l elect ion of the follow ing year, Higgi ns wa s elect ed to Dil ire ann for the first tim e.2002–200 7 .At the 2002 gener al election , Joe Hig gins retain ed his Dublin W est seat in Dil i reann. Clare Daly na rrowly missed out on gai ning a second seat for the party i n the Dublin Nor th cons tituen cy.Th e Anti-Bin Tax Camp aign ca me abou t at this ti me. On 19 Septem ber 2003, Higg ins a | |
| | 2 | Which specific parts of the text are colored red or blue? | The phrase 'l elect' is red. The segments 'e.2002-200' and 'th cons tituen cy.Th e Anti-Bin Tax Camp aign ca me abou' are blue. | |
| | 3 | What is visible in the background behind the paper? | A blurry background featuring a brick wall and a blue object at the top. | |
|
|
|
|
| ### Example 6 — 3 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | For documentation purposes, I need the text within this image transcribed. Could you provide that? | couples living together; 13.1% had a female householder with no husband present; and 31.0% were non-families. Of all households, 25.6% were made up of individuals, and 2.9% had someone living alone who was 65 years of age or older. The average household size was 2.59 and the average f | |
| | 2 | How is the text arranged on the page? | The text is arranged in two separate blocks or paragraphs, with a significant vertical gap of white space between the top block and the bottom block. | |
| | 3 | What font style is used for the text? | The text is printed in a serif font. | |
|
|
|
|
| ### Example 7 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Could you extract and type out the text from this image? | wealth of ot her su bjects too n um erous to list.Artists, designers, engraver s an d ad minist | |
| | 2 | Describe the people visible in the image. | Two people seen from behind: a woman with blonde hair wearing sunglasses on the left, and a man with dark hair on the right. | |
| | 3 | What is the background setting behind the people? | A garden with lush green foliage and pink flowers. | |
| | 4 | What colors are used for the text on the white paper? | The text appears in black, purple, and blue. | |
|
|
|
|
| ### Example 8 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | There's text embedded in this image that I need extracted. Can you transcribe it verbatim? | and society. Its mission includes "developing concepts of risk governance anticipating major risk issues an d providing risk governan ce policy advice for key d | |
| | 2 | What object is the text printed on? | A white sheet of paper. | |
| | 3 | How is the text arranged on the page? | The text is divided into two separate blocks with a large vertical gap between them. | |
| | 4 | What colors are visible in the background? | The background is out of focus but shows patches of blue and pink. | |
|
|
|
|
| ### Example 9 — 3 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | What messages or text are displayed in this image? I need them in a readable, typed format. | ir blowing through them which are us ed to replicate t he interaction bet ween air and an objec t flying through the a ir or moving al ong the groun d. Re searchers use win d tunnels | |
| | 2 | What object is the text printed on? | A white piece of paper. | |
| | 3 | What animal is visible in the background behind the paper? | A bird with yellow and blue feathers perched on a branch. | |
|
|
|
|
| ### Example 10 — 3 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | I need the written content from this picture converted into text. Can you do that? | re made up o f individuals, and 19.0 % had someone living al one who was 65 years of age or old er. The average h ousehold siz e was 2.67 and the ave rage family size was 3. 46.In the CDP th e population was spre ad out with 39.3 % under the age of 18 1 .8% from 18 to 2 4 33.9% fr om 25 to 44 16. 1% from 45 to 64 and 8. 9% who were 65 yea rs of age or older. The media n age was 34 years. For e very 100 females there were 93.1 males. For every 100 females age 18 and over there were 78.9 males. The median inc | |
| | 2 | What colors are used for the text in the image? | The text is primarily black, but specific numbers (19.0, 2.67, 3.46, 39.3, 1.8, 33.9, 16.1, 8.9, 34, 93.1, 78.9) and the acronym 'CDP' are printed in blue or purple. | |
| | 3 | How is the text arranged and formatted on the page? | The text is arranged in disjointed blocks on a white sheet of paper. The phrase 'rs of age or older.' is centered and printed in a larger font size than the surrounding text. | |
|
|
|
|
