Data and code for NeurIPS 2022 Paper "Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering".

For more details, please refer to the project page with dataset exploration and visualization tools: https://scienceqa.github.io.

🔔 If you have any questions or suggestions, please don't hesitate to let us know. You can directly email Pan Lu at UCLA using the email address [email protected], comment on the Twitter, or post an issue on this repository.

• [2023.05.04] ScienceQA Featured in Leaked Google Document: "We Have No Moat, And Neither Does OpenAI": A recent leak of an internal Google document highlights the advancements and impact of ScienceQA within the AI research community. 🎯
• [2023.05.03] In April, our ScienceQA dataset was downloaded 1,421 times from HuggingFace Datasets, showcasing its growing popularity in the community. [Link] 🌟
• [2023.04.19] Chameleon: Developed by UCLA and Microsoft, this innovative project achieves a new SOTA in the few-shot setting, reaching an impressive 86.54%. ⭐
• [2023.04.17] LLaVA: A collaborative effort by UW–Madison and Microsoft, this groundbreaking work sets a new SOTA at 92.53%. ⭐
• [2023.04.01] Our work is accepted by CVPR 2023 O-DRUM Workshop.
• [2023.04.01] Our work is covered by Towards AI.
• [2023.04.01] Our ScienceQA dataset was downloaded 377 times in March at HuggingFace Datasets.
• [2023.03.30] The ScienceQA dataset is now included at OpenDataLab.
• [2023.03.28] The ScienceQA dataset has served as the primary benchmark for LLaMA-Adapter, developed by Shanghai AI Laboratory, UCLA, and CUHK. ⭐
• [2023.02.13] Our work gives an oral presentation by Pan Lu at AAAI 2023 KnowledgeNLP Workshop.
• [2023.02.05] Our work is covered by MarkTechPost.
• [2023.02.24] The ScienceQA dataset is now included at HuggingFace Datasets. ⭐
• [2023.02.02] The ScienceQA dataset has served as the primary benchmark for the new generation of multimodal reasoning systems, Multimodal-CoT, developed by Amazon Science.
• [2022.11.29] Our work gives an poster presentation by Pan Lu at NeurIPS 2022.
• [2022.11.20] Our work is covered by Geek Culture | Medium.
• [2022.11] Our work is now included at Paper with Code.
• [2022.09.22] Our work is accepted to NeurIPS 2022. 🌟
• [2022.09.20] Our work is featured in Deep AI.

🔔 The leaderboard is continuously being updated. If you have any new results to contribute, please feel free to reach out to us.

Some notations in the table

• #Size: the model size (number of model parameters)
• #Param: number of tuned model parameters
• GPT-3: the text-davinci-002 engine
• GPT-4: the gpt-4 engine
• AE: the output is the answer followed by the explanation
• ALE: the output is the answer followed by the lecture and explanation
• T: the input only involves textual features
• Natural, Social, and Language Sciences: Questions are categorized into three subject areas: NAT (Natural Science), SOC (Social Science), and LAN (Language Science)
• Context Types: Questions are organized based on the type of context they include: TXT (Textual), IMG (Image), and NO (No Context)
• Grade Levels: Questions are divided into two groups based on the targeted grade levels: G1-6 (Grades 1-6) and G7-12 (Grades 7-12)

We present Science Question Answering (ScienceQA), a new benchmark that consists of 21,208 multimodal multiple choice questions with a diverse set of science topics and annotations of their answers with corresponding lectures and explanations. The lecture and explanation provide general external knowledge and specific reasons, respectively, for arriving at the correct answer.

ScienceQA, in contrast to previous datasets, has richer domain diversity from three subjects: natural science, language science, and social science. ScienceQA features 26 topics, 127 categories, and 379 skills that cover a wide range of domains.

We further design language models to learn to generate lectures and explanations as the chain of thought (CoT) to mimic the multi-hop reasoning process when answering ScienceQA questions. ScienceQA demonstrates the utility of CoT in language models, as CoT improves the question answering performance by 1.20% in few-shot GPT-3 and 3.99% in fine-tuned UnifiedQA.

For more details, you can find our project page here and our paper here.

The text part of the ScienceQA dataset is provided in data/scienceqa/problems.json. You can download the image data of ScienceQA by running:

. tools/download.sh

Alternatively, you can download ScienceQA from Google Drive and unzip the images under root_dir/data.

💥 The ScienceQA dataset is now available at HuggingFace Datasets!

For more details, you can explore the datatset and check the visualizations here: Explore and Visualizations.

python==3.8.10
huggingface-hub
nltk==3.5
numpy==1.23.2
openai==0.23.0
pandas==1.4.3
rouge==1.0.1
sentence-transformers==2.2.2
torch==1.12.1+cu113
transformers==4.21.1

Install all required python dependencies:

pip install -r requirements.txt

We use the image captioning model to generate the text content for images in ScienceQA. The pre-generated image captions are provided in data/captions.json.

(Optionally) You can generate the image captions with user-specific arguments with the following command, which will save the caption data in data/captions_user.json.

cd tools
python generate_caption.py

We build a few-shot GPT-3 model via chain-of-thought (CoT) prompting to generate the answer followed by the lecture and the explanation (QCM→ALE). The prompt instruction encoding for the test example in GPT-3 (CoT) is defined as below:

In our final model, we develop GPT-3 (CoT) prompted with two in-context examples and evalute it on the ScienceQA test split:

cd models
python run_gpt3.py \
--label exp1 \
--test_split test \
--test_number -1 \
--shot_number 2 \
--prompt_format QCM-ALE \
--seed 3

Our final GPT-3 (CoT) model achieves a state-of-the-art accuracy of 75.17% on the test split. One prediction example is visualized below. We can see that GPT-3 (CoT) predicts the correct answer and generates a reasonable lecture and explanation to mimic the human thought process.

We can get the accuracy metrics on average and across different question classes by running:

cd tools
python evaluate_acc.py

We can run the following command to evaluate the generated lectures and explanations automatically:

cd tools
python evaluate_explaination.py

You can try other prompt templates. For example, if you want the model to take the question, the context, and the multiple options as input, and output the answer after the lecture and explanation (QCM→LEA), you can run the following script:

cd models
python run_gpt3.py \
--label exp1 \
--test_split test \
--test_number -1 \
--shot_number 2 \
--prompt_format QCM-LEA \
--seed 3

• 🔥 Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering, NeurIPS 2022 [ArXiv] [Project] [Code] [Hugging Face Dataset]
• 🔥 Multimodal Chain-of-Thought Reasoning in Language Models, arXiv:2302.00923 [ArXiv] [Code]
• 🔥 LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention, arXiv:2303.16199 [ArXiv] [Code] [Gradio Demo] [LightingAI Report]
• 🔥 Visual Instruction Tuning, arXiv:2304.08485 [ArXiv] [Project] [Code] [Demo] [Dataset]
• 🔥 Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models, arXiv:2304.09842 [ArXiv] [Project] [Code] [YouTube]
• 🔥 LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model, arXiv:2304.15010 [ArXiv] [Code] [YouTube]

This work is licensed under a MIT License.

The ScienceQA dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

If the paper, codes, or the dataset inspire you, please kindly cite us:

@inproceedings{lu2022learn,
    title={Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering},
    author={Lu, Pan and Mishra, Swaroop and Xia, Tony and Qiu, Liang and Chang, Kai-Wei and Zhu, Song-Chun and Tafjord, Oyvind and Clark, Peter and Ashwin Kalyan},
    booktitle={The 36th Conference on Neural Information Processing Systems (NeurIPS)},
    year={2022}
}