leinms/flickr30k-qwen3vl-baseline

Flickr30k Qwen3-VL Baseline Captions (Test Split)

This dataset is based on the Mozilla/flickr30k-transformed-captions-gpt4o test split and contains 1,000 images from the original Flickr30k dataset. It includes both the original metadata and newly generated baseline captions produced using the Qwen3-VL-2B-Instruct vision-language model.

Contents

Each entry includes:

• —image — the original Flickr30k image
• —alt_text — GPT-4o transformed caption from Mozilla's de-biasing pipeline
• —original_alt_text — list of 5 human-written captions
• —objects — detected objects metadata
• —baseline_answ — new caption generated by Qwen3-VL-2B-Instruct

Generation Settings

All baseline captions in baseline_answ were generated with:

• —Model: Qwen/Qwen3-VL-2B-Instruct
• —Precision: bfloat16
• —Attention: flashattention2
• —Device: CUDA
• —Sampling: (do_sample = True)
• —Max new tokens: 64

Prompt Used

Each image was processed with the following prompt:

"Describe this image in one concise English sentence."

Purpose

This dataset is intended for:

• —benchmarking vision-language models,
• —evaluating caption quality across systems,
• —studying differences between GPT-4o, human captions, and Qwen-generated captions.

The dataset contains only the test subset, making it lightweight and easy to experiment with.