This challenge uses retail product records from several product groups. The information in each record varies. Some records have detailed source text, while others have missing or masked values.

Your task is to enrich each record in Test_Masked.csv with accurate, useful product content in the ten output fields defined in ``submission_format.json``. The aim is to create clear product information that describes the exact item. This is not a request to fill every field. Every value should be supported by the supplied information or reliable evidence allowed by the content guide.

Start by confirming the product identity from its name, brand, category, descriptions, and specifications. Pay close attention to pack size, flavour, shade, weight, dimensions, and units. Similar products can differ in important ways, so do not transfer a detail from another item unless you can confirm it applies to the exact product.

Use unlabelled_data.csv to understand the catalogue and develop an approach that works across products. It has no target answers and should not be treated as a verified answer source for a test record.

Treat product identifiers as text. Preserve every record_id exactly and copy it into article_id in your submission.

Some test inputs contain missing or masked information. If a detail cannot be verified, leave its output cell blank. If a field does not apply, follow do_not_predict_fields and leave its output cell blank. If sources conflict, do not choose a value without supporting evidence. Use submission_format.json for the required columns and empty value shapes.

Phase 1 Constraints

• End to end inference time: 3 hours maximum
• All in inference cost: US$20 maximum
• GPU memory: One GPU with at most 24 GB VRAM
• Retries: At most one retry for a failed call, included in the time and cost caps