Sparkles: resources for SparklesChat, SparklesDialogue and SparklesEval

Sparkles: Unlocking Chats Across Multiple Images for Multimodal Instruction-Following Models — Yupan Huang, Zaiqiao Meng, Fangyu Liu, Yixuan Su, Nigel Collier, Yutong Lu. ICLR 2024 Workshop on Data Problems for Foundation Models (DPFM).

Code: HYPJUDY/Sparkles · Demo video: YouTube

This repository is a complete, unmodified mirror of the resource folder that was previously distributed through a OneDrive link. Its layout is exactly the Sparkles/ root expected by the code (dataset/config_path.py, train_configs/sparkleschat.yaml, eval_configs/sparkles_eval.yaml):

pip install -U "huggingface_hub[hf_xet]"
hf download HYPJUDY/Sparkles --repo-type dataset --local-dir /path/to/Sparkles
# then set sparkles_root="/path/to/Sparkles/" in dataset/config_path.py and the yaml configs

Contents

• models/pretrained/sparkleschat_7b.pth — SparklesChat (7B) weights. Like MiniGPT-4, only the projection layer is trained and saved (38 MB); you also need Vicuna-7B v0 and the BLIP-2 vision encoder, see the code README. minigpt4_7b_stage1.pth is the MiniGPT-4 first-stage checkpoint used as the starting point (redistributed from Vision-CAIR/MiniGPT-4).
• data/SparklesDialogueCC/ — 4,521 GPT-4-generated multi-image dialogues over 3,373 Conceptual Captions images (annotations/*.json, HTML visualisation, images/).
• data/SparklesDialogueVG/ — 2,000 dialogues over 7,000 Visual Genome images.
• data/SVIT/, data/LLaVA/ — the SVIT and LLaVA detailed descriptions used as caption sources, both the original files and the *_filtered_for_Sparkles.json versions converted to the Sparkles format.
• data/cc_sbu_align/, data/VisualGenome/image_data.json — the MiniGPT-4 stage-2 alignment data and the Visual Genome image index, as used by the data-generation scripts.
• evaluation/ — SparklesEval (150 dialogues, 550 Visual Genome images), plus the BISON and NLVR2 subsets used in the paper, each with annotations/, images/ and the results/ of SparklesChat and MiniGPT-4 (7B).
• assets/ — images used in the paper's examples and the dataset statistics figures.
• data/*.zip, evaluation.zip, models.zip — convenience archives of the folders above with identical contents.

Sparkles
├── models/pretrained/{sparkleschat_7b.pth, minigpt4_7b_stage1.pth}
├── data
│   ├── SparklesDialogueCC/{annotations, images}
│   ├── SparklesDialogueVG/{annotations, images}
│   ├── cc_sbu_align/{image, filter_cap.json, metadata.jsonl}
│   ├── SVIT/{detail_description.json, SVIT_detail_description_filtered_for_Sparkles.json}
│   ├── LLaVA/{complex_reasoning_77k.json, detail_23k.json, *_filtered_for_Sparkles.json}
│   └── VisualGenome/image_data.json
├── evaluation
│   ├── SparklesEval/{annotations, images, results}
│   ├── BISON/{annotations, images, results}
│   └── NLVR2/{annotations, images, results}
└── assets/{images, statistics}

License and third-party data

Our code is distributed under the BSD 3-Clause License, our model is subject to LLaMA's model license, and the dialogues were generated with GPT-4 and are subject to OpenAI's Terms of Use. Third-party data is redistributed unchanged for reproducibility and keeps its original terms: Visual Genome (CC BY 4.0), Conceptual Captions images via MiniGPT-4 cc_sbu_align, SVIT (CC BY 4.0), LLaVA-Instruct-150K (CC BY 4.0), BISON (CC BY-NC 4.0) with MS-COCO images, NLVR2 (annotations CC BY 4.0; the image owners hold the copyright of the images, which are included here for research use only), and the MiniGPT-4 checkpoint (BSD 3-Clause).

Citation

@article{huang2023sparkles,
  title={Sparkles: Unlocking Chats Across Multiple Images for Multimodal Instruction-Following Models},
  author={Huang, Yupan and Meng, Zaiqiao and Liu, Fangyu and Su, Yixuan and Nigel, Collier and Lu, Yutong},
  journal={arXiv preprint arXiv:2308.16463},
  year={2023}
}