Datasets:
- How this was cleaned
answervsorg_answer- Schema
- What is NOT claimed
- Examples
- Example 1 — 29 turns on one image
- Example 2 — 5 turns on one image
- Example 3 — 5 turns on one image
- Example 4 — 3 turns on one image
- Example 5 — 3 turns on one image
- Example 6 — 4 turns on one image
- Example 7 — 4 turns on one image
- Example 8 — 5 turns on one image
- Example 9 — 7 turns on one image
- Example 10 — 8 turns on one image
- Example 1 — 29 turns on one image
cocotext_train_cleaned
The cocotext_train family of the ElliotVL supervised-fine-tuning pool, after VLM cleaning.
| images | 41,200 |
| QA turns | 238,897 |
| answers rewritten by the cleaning pass | 0 |
QA created by the cleaning pass (new_qa) |
not measured for this family |
| shards | 21 |
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_answeris empty throughout: recovering it means scanning the whole pre-clean family, which is raw family is 10 GB (> 6 GB cap). The cleanedansweris 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— HFImage(), renders directly in the dataset viewerimage_sha256— content hash of the imagen_qa— number of turns attached to this imageqa— list of {question,answer,org_answer}, in source orderfamily/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.
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 — 29 turns on one image
| # | question | answer |
|---|---|---|
| 1 | Extract all visible text content from this image. | GRME |
| 2 | What text is visible on the woman's black tank top? | ALL-STAR GAME |
| 3 | What is the man in the center wearing? | A red short-sleeved shirt, black pants, black shoes, and a yellow wristband. |
| 4 | Describe the baseball player's uniform and equipment. | He wears a white uniform with vertical pinstripes, a black batting helmet, white batting gloves, and holds a light-colored wooden bat. |
| 5 | Where are the extra baseball bats and batting gloves located? | They are lying on the black on-deck circle mat behind the player. |
| 6 | What are the colors of the two main ground surfaces in the image? | The dirt is brown and the grass is green. |
| 7 | How many people are visible and how are they arranged? | Three people: a woman on the left, a man in the center, and a baseball player on the right near the on-deck circle. |
| 8 | What is the woman on the left doing? | She is standing on the dirt, smiling, and facing the man in the red shirt. |
| 9 | What is the man in the red shirt holding? | A small black device in his right hand. |
| 10 | What color is the wristband on the man in the red shirt's wrist? | Yellow. |
| 11 | What piece of jewelry is visible on the woman? | A necklace. |
| 12 | What are the colors of the on-deck circle mat and the logo on it? | The mat is black with a yellow logo/design. |
| 13 | How many baseball bats are lying on the on-deck circle mat? | Two. |
| 14 | What color are the baseball player's batting gloves? | White. |
| 15 | What is the baseball player doing with his left hand? | He is extending his left hand outward with an open palm. |
| 16 | What color are the woman's boots? | Black. |
| 17 | Which hand is the baseball player using to hold the bat? | His right hand. |
| 18 | What is the total number of baseball bats visible in the image? | Three — one held by the player and two lying on the on-deck circle mat. |
| 19 | What color are the batting gloves lying on the on-deck circle mat? | Red and white. |
| 20 | What color is the baseball player's helmet? | Black. |
| 21 | What is clipped to the waist of the man in the red shirt? | An ID badge or credential. |
| 22 | What is the woman on the left holding in her right hand? | A small black object, possibly a phone or microphone. |
| 23 | What is on the baseball player's left wrist? | A white wristband or athletic tape. |
| 24 | Where is the on-deck circle mat located relative to the man in the red shirt? | To his right. |
| 25 | Which way is the baseball player looking? | Toward the woman and the man in the red shirt. |
| 26 | What is the shape of the on-deck circle mat? | Circular. |
| 27 | What color is the baseball bat the player is holding? | Light tan or natural wood. |
| 28 | From what general perspective was this photo taken? | From an elevated angle, likely from the stands. |
| 29 | What is visible on the baseball player's left sleeve? | A patch or emblem. |
Example 2 — 5 turns on one image
| # | question | answer |
|---|---|---|
| 1 | Read and extract the text from this image. | PowerShot, 2013 |
| 2 | What year is written on the headband worn by the woman in the center? | 2013 |
| 3 | Describe the person standing to the left of the woman with the camera. | A person with red-streaked hair wearing a purple jacket with a fur-lined hood, holding up a smartphone. |
| 4 | What is the main subject in the foreground holding? | A black DSLR camera with a strap that says 'PowerShot'. |
| 5 | What background elements suggest a public event or celebration? | A crowd of people, a bus in the distance, and balloons (yellow and green) visible in the top right. |
Example 3 — 5 turns on one image
| # | question | answer |
|---|---|---|
| 1 | What is the woman wearing? | A white t-shirt and a blue denim skirt. |
| 2 | What vehicle is parked next to the parking meter? | A white SUV. |
| 3 | What number is visible on the green sticker on the meter? | 4592 |
| 4 | What is the woman doing? | She is pressing a button on the parking meter keypad. |
| 5 | What type of trees are visible in the background? | Palm trees. |
Example 4 — 3 turns on one image
| # | question | answer |
|---|---|---|
| 1 | What specific internet slang terms are written on the whiteboard behind the man? | The whiteboard lists terms including 'o_O', 'pwn3d', 'ROFL', ':D', 'NOT', 'h4x0r', 'ub3r', 'g33k', 'n00b', and 'meh'. |
| 2 | What text is printed on the man's red t-shirt? | The text on the shirt reads 'No. I will not fix computer.' |
| 3 | What is the man doing in the image? | The man is holding a piece of food, which looks like a bagel or bread, up to his mouth as if taking a bite. |
Example 5 — 3 turns on one image
| # | question | answer |
|---|---|---|
| 1 | What animal is standing on the left side of the image? | A dark tabby cat. |
| 2 | What object is sitting on the floor to the right of the bottle? | A silver metal pet bowl on a black rubber base. |
| 3 | What is visible through the window in the background? | Wooden kitchen cabinets and a hanging white towel. |
Example 6 — 4 turns on one image
| # | question | answer |
|---|---|---|
| 1 | Read all the text in this image. | © Earth Touch |
| 2 | What two animals are visible in the image? | An adult elephant and a baby elephant. |
| 3 | What is the baby elephant doing? | Walking through the grass near the adult elephant. |
| 4 | What distinct white feature is visible on the adult elephant? | A tusk. |
Example 7 — 4 turns on one image
| # | question | answer |
|---|---|---|
| 1 | Identify the text contained in this image. | K5 KING5.com |
| 2 | What is the woman in the foreground doing? | She is sitting at a desk working on a computer. |
| 3 | What items are visible on the metal shelving unit to the left of the woman? | There are stacks of books, papers, and a small potted plant. |
| 4 | Describe the office environment shown in the background. | It is an open-plan office with beige cubicle walls, desks, and large windows letting in daylight. |
Example 8 — 5 turns on one image
| # | question | answer |
|---|---|---|
| 1 | What is the brand name and flavor written on the bottle label? | Sleeping Giant Fruit Winery, Blackberry |
| 2 | What are the volume and alcohol content listed at the bottom of the label? | 750 ml and 11.0% alc./Vol. |
| 3 | Where is the product from according to the label? | British Columbia, Canada |
| 4 | Describe the object to the right of the bottle. | A clear wine glass containing a small amount of red liquid. |
| 5 | What surface are the bottle and glass resting on? | A speckled brown countertop. |
Example 9 — 7 turns on one image
| # | question | answer |
|---|---|---|
| 1 | Transcribe all legible text in the image. Return a JSON list of objects with keys text and bbox, where bbox is [x1, y1, x2, y2] normalized to [0, 1]. Use an empty list if there is no legible text. |
[] |
| 2 | Why does the Siamese cat have pointed ears? | The breed standard for Siamese cats specifies that they should have pointed (triangular) ears as part of their physical characteristics. This trait helps distinguish them from other breeds like Persian or British Shorthair which typically have rounded ears. Pointed ears can also be associated with cooler temperaments according to some theories within feline genetics. |
| 3 | What emotion can the viewer feel upon seeing this cat sitting calmly inside? | The viewer might feel calm or peaceful when looking at this serene-looking Siamese cat sitting quietly in its wooden enclosure. The gentle expression of the feline could evoke feelings of tranquility as it appears content with its surroundings. This stillness may also inspire viewers to appreciate moments of quiet reflection within their own lives. Ultimately, emotions experienced by individuals will vary based on personal experiences and emotional responses to pets and animals. However, overall, one is likely to experience positive emotions such as comfort, relaxation, or affection towards the cute appearance of the cat. |
| 4 | What color collar does the cat have? | The cat has a black collar with a bell attached. |
| 5 | Why are the cat's eyes glowing bright white? | The eyes are glowing due to the camera flash reflecting off the tapetum lucidum, a common effect in flash photography of animals. |
| 6 | What object is visible on the left side of the wooden structure? | There is a mirror on the left side, which reflects the back of the cat. |
| 7 | Describe the texture and color of the walls behind the cat. | The walls are made of light-colored wooden panels with visible knots, resembling knotty pine. |
Example 10 — 8 turns on one image
| # | question | answer |
|---|---|---|
| 1 | Transcribe all legible text in the image. Return a JSON list of objects with keys text and bbox, where bbox is [x1, y1, x2, y2] normalized to [0, 1]. Use an empty list if there is no legible text. |
[] |
| 2 | What is the main piece of furniture in the foreground? | A bed with white bedding and pillows. |
| 3 | What is mounted on the wall above the head of the bed? | A wall sconce lamp. |
| 4 | What large wooden object is visible next to the bed? | A closed wooden door. |
| 5 | What furniture is located on the right side of the room? | A wooden dresser or vanity table with a round mirror and a lamp. |
| 6 | What is hanging on the back of the chair? | A white bag. |
| 7 | What small appliance is visible next to the dresser? | A small white refrigerator or mini-fridge. |
| 8 | What type of flooring is visible in the room? | Hardwood floor. |
- Downloads last month
- 26









