ZovdaPaired records
One image, one title, one attribute set
One product image with its title and attribute set, kept as a single record.
Datuva by Zovda · Training data
Image–text pairs with evidence, not guesses.
E-commerce image–text pairs and attribute-extraction records where each output field points back to the text that supports it. The current release holds 2,808 records in two forms: full text and title only.
- 2,808 records
- Field-level evidence
- Two input forms
Collection contents
What ships in this collection.
Evidence chains
Every field traced to raw text
Each extracted value maps to the raw text that supports it. Values without support are dropped and recorded with a reason.
Two input forms
Full text and title only
The same task ships in two variants so you can train on different input budgets.
Field-level JSON
No parsing prose for labels
Instruction, input, output, evidence, and meta are separate fields — no parsing prose to get the label.
Record structure
What each record holds.
| Field | What it holds |
|---|---|
instruction | The extraction task given to the model. |
input | Product image reference and the text the model may read. |
output | Structured attributes for the record. |
evidence | Raw span → normalized value per field, plus a drop reason when a value is rejected. |
meta | Form (full text or title only), version, and record identifiers. |
Training stage
Recommended use by stage.
| Stage | How it is used |
|---|---|
| Alignment | Image–text pairs for multimodal alignment. |
| SFT | Extraction records with evidence chains. |
| Evaluation | Cases with and without supporting text, for boundary checks. |
Delivery
One structure per collection.
| Layer | Contents |
|---|---|
| manifest.json | Every file with name, size, date, and collection ID. |
| UTF-8 JSONL | One record per line for text and annotations. |
| Documents | Data card, field schema, and a PDF spec for each collection. |
| UTF-8 JSONL | One record per line; image references resolve through the manifest. |
Request access
Request the image–text catalog.
Tell us the category, input form, and volume. We reply within one business day. Sample packs, data cards, and field schemas are shared after your request is reviewed.
FAQ
Before you request a collection.
What does evidence mean here?
For each output field the record stores the raw text span that supports it, or the reason it was dropped. That makes over-inference visible during training and review.
What are the two forms?
Full text and title only. Same task, different amount of available context.
Can we take one product category?
Yes. Records can be sliced by category, and the schema can be adapted before production.