Banana ViDoRe v3 Fullpipe

Nano Banana Pro full-pipeline synthetic training data for ViDoRe v3 finance and industrial domains.

This repository contains 4670 training records and 40190 unique referenced images across domain-separated ColFlor/ColQwen training splits. Images are included in the repository and paths in each JSONL are relative to that domain directory.

Generated at: 2026-06-29T09:39:01.644860+00:00

Layout

finance/train.jsonl
finance/metadata.json
finance/images/...
industrial/train.jsonl
industrial/metadata.json
industrial/images/...
metadata.json

Splits

Usage

from banana.training.data.dataset import BananaDataset

finance = BananaDataset("finance/train.jsonl", num_hard_negatives=4)
industrial = BananaDataset("industrial/train.jsonl", num_hard_negatives=4)

When loading from the Hub with the Banana training scripts, use the domain JSONL path:

python -m banana.training.scripts.train \
  --data-path hf://datasets/vkehfdl1/banana-vidorev3-fullpipe/finance/train.jsonl

Schema

Each JSONL row contains:

• query: synthetic training query
• positive_pages: relative image paths under the same domain directory
• hard_negative_pages: relative hard-negative image paths
• metadata.domain: finance or industrial
• metadata.source_dataset: source ViDoRe v3 dataset name
• metadata.negative_method: banana_fullpipe

The data was generated for training/adaptation only. Use held-out ViDoRe test splits for evaluation.