| --- |
| license: other |
| task_categories: |
| - visual-question-answering |
| language: |
| - en |
| --- |
| |
| # cord_train_cleaned |
|
|
| The **cord_train** family of the ElliotVL supervised-fine-tuning pool, **after VLM cleaning**. |
| |
| | | | |
| |---|---| |
| | images | 794 | |
| | QA turns | 2,990 | |
| | answers rewritten by the cleaning pass | 303 | |
| | QA created by the cleaning pass (`new_qa`) | 2,197 (73.5%) | |
| | shards | 4 | |
| |
| ## 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. |
|
|
| 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 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":[{"cnt":"x 1","nm":"Nasi Campur Bali","price":"75,000"},{"cnt":"x 1","nm":"Bbk Bengil Nasi","price":"125,000"},{"cnt":"x 1","nm":"MilkShake Starwb","price":"37,000"},{"cnt":"x 1","nm":"Ice Lemon Tea","price":"24,000"},{"cnt":"x 1","nm":"Nasi Ayam Dewata","price":"70,000"},{"cnt":"x 3","nm":"Free Ice Tea","price":"0"},{"cnt":"x 1","nm":"Organic Green Sa","price":"65,000"},{"cnt":"x 1","nm":"Ice Tea","price":"18,000"},{"cnt":"x 1","nm":"Ice Orange","price":"29,000"},{"cnt":"x 1","nm":"Ayam Suir Bali","price":"85,000"},{"cnt":"x 2","nm":"Tahu Goreng","price":"36,000"},{"cnt":"x 2","nm":"Tempe Goreng","price":"36,000"},{"cnt":"x 1","nm":"Tahu Telor Asin","price":"40,000."},{"cnt":"x 1","nm":"Nasi Goreng Samb","price":"70,000"},{"cnt":"x 3","nm":"Bbk Panggang Sam","price":"366,000"},{"cnt":"x 1","nm":"Ayam Sambal Hija","price":"92,000"},{"cnt":"x 2","nm":"Hot Tea","price":"44,000"},{"cnt":"x 1","nm":"Ice Kopi","price":"32,000"},{"cnt":"x 1","nm":"Tahu Telor Asin","price":"40,000"},{"cnt":"x 1","nm":"Free Ice Tea","price":"0"},{"cnt":"x 1","nm":"Bebek Street","price":"44,000"},{"cnt":"x 1","nm":"Ice Tea Tawar","price":"18,000"}],"sub_total":{"etc":"-45","service_price":"100,950","subtotal_price":"1,346,000","tax_price":"144,695"},"total":{"total_price":"1,591,600"}} | |
| | 2 | What surface is the receipt placed on? | A black and white woven placemat. | |
| | 3 | What specific text labels appear in the calculation section below the item list? | Sub-Total, Service, PBI, Rounding, Grand Total. | |
| | 4 | Is the top header of the receipt visible? | No, the top section containing the header information is blurred out. | |
| |
| |
| ### Example 2 — 4 turns on one image |
| |
|  |
| |
| | # | question | answer | |
| |---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":[{"cnt":"1","nm":"SPGTHY BOLOGNASE","price":"58,000"},{"cnt":"1","nm":"PEPPER AUS","price":"165,000","sub_nm":"WELL DONE"},{"cnt":"1","nm":"WAGYU RIBEYE","price":"195,000","sub_nm":"MEDIUM WELL"},{"cnt":"1","nm":"ICED LEMON TEA","price":"22,000"},{"cnt":"1","nm":"FUSION TEA LYCHE","price":"28,000"},{"cnt":"1","nm":"NUTTELA BROWNIES","price":"35,000"}],"sub_total":{"service_price":"25,150","subtotal_price":"503,000","tax_price":"52,815"},"total":{"total_price":"580,965"}} | |
| | 2 | What table number is listed on the receipt? | 6 / 1 | |
| | 3 | What specific abbreviations are used for the tax and service charge lines? | PB1 and SVC CHRG | |
| | 4 | What pattern is visible on the surface underneath the receipt? | A plaid or checkered pattern | |
|
|
|
|
| ### Example 3 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":[{"cnt":"4","nm":"HAKAU UDANG","price":"92,000"},{"cnt":"4","nm":"SIAO MAI BABI","price":"80,000"},{"cnt":"3","nm":"CEKER AYAM","price":"60,000"},{"cnt":"2","nm":"BAKPAO BKR C CRISPY","price":"42,000"},{"cnt":"3","nm":"TAHU GORENG CRISPY","price":"60,000"}],"sub_total":{"subtotal_price":"334,000"},"total":{"cashprice":"350,000","changeprice":"-16,000","menuqty_cnt":"16","menutype_cnt":"5","total_price":"334,000"}} | |
| | 2 | Which labels in the totals section are printed in bold text? | Total, CASH, and Cash CHANGE | |
| | 3 | What is the status of the text at the very top and very bottom of the receipt? | The text is blurred out and illegible. | |
| | 4 | For which items is the word 'CRISPY' placed on a separate line below the main name? | BAKPAO BKR C and TAHU GORENG | |
| |
| |
| ### Example 4 — 4 turns on one image |
| |
|  |
| |
| | # | question | answer | |
| |---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":[{"cnt":"1","nm":"Coke (L)","price":"25.000"},{"cnt":"1","nm":"Extra Jelly Lychee","price":"5.000"},{"cnt":"1","nm":"Popcorn Salt (M)","price":"40.000"}],"total":{"cashprice":"70.000","changeprice":"0","total_price":"70.000"}} | |
| | 2 | What is the name of the cafe printed in the logo at the top of the receipt? | XXI café | |
| | 3 | What are the specific labels used for the payment amount and the change returned? | CASH and CHANGED | |
| | 4 | Describe the physical context of how the receipt is being presented. | The receipt is being held by a person's fingers, visible at the top left and bottom right corners. | |
|
|
|
|
| ### Example 5 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":{"cnt":"x1","nm":"CHEESE JOHN","num":"2005","price":"9.500,00","unitprice":"9.500,00"},"total":{"cashprice":"20.000,00","changeprice":"10.500,00","menuqty_cnt":"1","total_price":"9.500,00"}} | |
| | 2 | What specific text labels appear in the payment summary section at the bottom of the receipt? | The labels are 'Total.', 'Cash Tendered:', and 'Change:'. | |
| | 3 | Describe the visual imperfections visible on the receipt paper near the middle. | There are dark ink smudges or stains obscuring part of the area near the 'Total Items' line. | |
| | 4 | What object is visible at the very bottom edge of the image, below the receipt? | A metal clip or latch mechanism is visible. | |
|
|
|
|
| ### Example 6 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":[{"cnt":"1","nm":"THAI ICED COFFEE","price":"20.000","unitprice":"@20.000"},{"cnt":"1","nm":"THAI ICED GREEN TEA","price":"22.000","unitprice":"@22.000"}],"sub_total":{"subtotal_price":"42.000"},"total":{"cashprice":"42.000","changeprice":"0","total_price":"42.000"}} | |
| | 2 | What is the exact text label used for the final total amount on the receipt? | GRANDTOTAL | |
| | 3 | What is the exact text label for the subtotal? | SUB-TOTAL | |
| | 4 | Describe the graphic visible in the top left corner of the receipt. | A black silhouette of an animal head, resembling a bull or elephant. | |
| |
| |
| ### Example 7 — 4 turns on one image |
| |
|  |
| |
| | # | question | answer | |
| |---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":{"cnt":"1x","nm":"Cheese Twist","price":"12.000","unitprice":"12.000"},"sub_total":{"subtotal_price":"12.000","tax_price":"1.200"},"total":{"emoneyprice":"13.200","total_price":"13.200"}} | |
| | 2 | What is the tax percentage listed on the receipt? | 10.0% | |
| | 3 | What two icons are visible at the bottom left of the receipt? | A chain link icon and a camera icon. | |
| | 4 | Describe the surface on which the receipt is placed. | A light-colored wooden surface with visible grain. | |
| |
| |
| ### Example 8 — 4 turns on one image |
| |
|  |
| |
| | # | question | answer (cleaned) | org_answer (pre-judge) | |
| |---|---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":{"cnt":"1","nm":"Classic Milktea Black [L]","price":"22,000"},"total":{"cashprice":"50,000","changeprice":"28,000","total_price":"22,000"}} | {"menu":{"cnt":"1","nm":"Classic Milktea","price":"22,000"},"total":{"cashprice":"50,000","changeprice":"28,000","total_price":"22,000"}} | |
| | 2 | What is the full name of the item purchased, including flavor and size? | Classic Milktea Black [L] | _(unchanged)_ | |
| | 3 | What is holding the receipt in the image? | A person's hand | _(unchanged)_ | |
| | 4 | What color object is visible in the top left corner of the image? | Blue fabric | _(unchanged)_ | |
|
|
|
|
| ### Example 9 — 3 turns on one image |
|
|
|  |
|
|
| | # | question | answer (cleaned) | org_answer (pre-judge) | |
| |---|---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"items": [{"name": "Kupon 13", "price": "49,091"}], "subtotal": "49,091", "tax": {"name": "PBI (10%)", "price": "4,909"}, "total": {"name": "Dine In Total", "price": "54,000"}, "payment": {"cash": "Rp. 100,000", "change": "46,000"}} | {"menu":{"nm":"Kupon 13","price":"49,091"},"sub_total":{"subtotal_price":"49,091","tax_price":"4,909"},"total":{"cashprice":"Rp. 100,000","changeprice":"46,000","total_price":"54,000"}} | |
| | 2 | What color is the receipt paper? | Pink | _(unchanged)_ | |
| | 3 | What is the status of the text at the top of the receipt? | It is blurred out. | _(unchanged)_ | |
|
|
|
|
| ### Example 10 — 4 turns on one image |
|
|
|  |
|
|
| | # | question | answer | |
| |---|---|---| |
| | 1 | Extract the structured receipt information from the image. Return only a JSON object with the receipt fields. | {"menu":{"nm":"NASI 2SAYUR TELR","price":"15000"},"total":{"cashprice":"15000","menuqty_cnt":"1Q"}} | |
| | 2 | What surface is the receipt placed on? | A light-colored wooden table. | |
| | 3 | What are the text labels for the quantity and payment method? | "ITEMS" and "CASH". | |
| | 4 | What is the legible text at the very bottom of the receipt? | "TERIMA KASIH". | |
| |
| |
