ScienceQA: Science Question Answering

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 lupantech@gmail.com, comment on the Twitter, or post an issue on this repository.

💥 News 💥

• [2023.12.29] 🚨 We have a major update featuring over 100 recent models! We appreciate your contributions and feedback. 🚀
• [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.

🌟 Star History

🔥 Leaderboard 🔥

Evaluation of different methods on the test split (whole: 4,241, mini: 1,000 examples). The accuracies across various categories and the overall average are reported below.

😀 You are invited to contribute your results to the TabMWP test split! Please send your result scores to this email or open a new issue at the github repository.

⚠️⚠️⚠️ Caveat: The data in the leaderboard is collected manually from existing papers. There might be some errors in the data, ambiguous data due to different interpretations, and missing data due to the lack of information in the papers. Make sure to double-check the data before using it. Please contact us at this email if you find any errors or have any suggestions. We appreciate your contributions and feedback.

The interactive leaderboard is available at https://scienceqa.github.io/leaderboard.html.

#	Model	Method	Learning	#Size	#P	Link	Date	NAT	SOC	LAN	TXT	IMG	NO	G1-6	G7-12	Avg
*	Human Performance	-	-	-	-	Link	22-09-20	90.23	84.97	87.48	89.60	87.50	88.10	91.59	82.42	88.40
*	Random Chance	-	-	-	-	Link	22-09-20	40.28	46.13	29.25	47.45	40.08	33.66	39.35	40.67	39.83
1	Mutimodal-T-SciQ_Large 🥇	LLM	Fine-tune	738M	738M	Link	23-05-05	96.89	95.16	95.55	96.53	94.70	96.79	96.44	95.72	96.18
2	MC-CoT_F-Large 🥈	VLM	Fine-tune	783M	-	Link	23-11-23	97.47	90.44	93.18	96.97	93.75	94.49	95.30	94.13	94.88
3	Honeybee (Vicuna-13B) 🥉	VLM	Fine-tune	13B	-	Link	23-12-11	95.20	96.29	91.18	94.48	93.75	93.17	95.04	93.21	94.39
4	Enigma-COT_Large	LLM	Fine-tune	793M	793M	Link	23-07-24	97.51	84.70	94.73	96.68	91.37	95.89	94.46	93.47	94.11
5	MC-CoT_Large	VLM	Fine-tune	738M	-	Link	23-11-23	95.47	89.99	91.82	95.11	92.66	93.24	94.27	91.76	93.37
6	DPMM-CoT_Large	VLM	Fine-tune	738M	738M	Link	23-12-14	95.52	90.33	91.36	95.50	93.26	92.68	93.28	93.47	93.35
7	LLaVA (GPT-4 judge)	VLM	Fine-tune	13B	13B	Link	23-04-17	91.56	96.74	91.09	90.62	88.99	93.52	92.73	92.16	92.53
8	CoMD (Vicuna-7B)	VLM	Fine-tune	7B	-	Link	23-11-14	91.83	95.95	88.91	90.91	89.94	91.08	92.47	90.97	91.94
9	Mutimodal-T-SciQ_Base	LLM	Fine-tune	223M	223M	Link	23-05-05	91.52	91.45	92.45	91.94	90.33	92.26	92.11	91.10	91.75
10	Multimodal-CoT_Large	VLM	Fine-tune	738M	738M	Link	23-02-02	95.91	82.00	90.82	95.26	88.80	92.89	92.44	90.31	91.68
11	PILL (LLaMA-7B)	VLM	Fine-tune	7B	45M	Link	23-11-03	90.36	95.84	89.27	89.39	88.65	91.71	92.11	89.65	91.23
12	LLaVA (ViT-L/16-224)	VLM	Fine-tune	13B	-	Link	23-12-04	-	-	-	-	-	-	-	-	91.2
13	DPMM-CoT_Base	VLM	Fine-tune	223M	223M	Link	23-12-14	92.72	87.85	89.91	92.72	90.48	91.29	91.45	90.11	90.97
14	LLaVA	VLM	Fine-tune	13B	13B	Link	23-04-17	90.36	95.95	88.00	89.49	88.00	90.66	90.93	90.90	90.92
15	LaVIN-13B	VLM	Fine-tune	13B	5.4M	Link	23-05-24	89.88	94.49	89.82	88.95	87.61	91.85	91.45	89.72	90.83
16	MC-CoT_F-Base	VLM	Fine-tune	248M	-	Link	23-11-23	93.56	83.58	90.73	94.13	89.24	90.94	90.93	90.38	90.73
17	MC-CoT_Base	VLM	Fine-tune	223M	-	Link	23-11-23	91.87	84.59	93.00	92.28	88.30	92.75	90.64	90.64	90.64
18	LLaMA-SciTune	VLM	Zero-shot	13B	-	Link	23-07-03	89.30	95.61	87.00	93.08	86.67	91.75	84.37	91.30	90.03
19	LaVIN-7B	VLM	Fine-tune	7B	3.8M	Link	23-05-24	89.25	94.94	85.24	88.51	87.46	88.08	90.16	88.07	89.41
20	Flan-T5-XL (LoRA)	LLM	Fine-tune	-	-	Link	23-11-03	-	-	-	-	-	-	-	-	89.29
21	Chat-UniVi	VLM	Fine-tune	7B	-	Link	23-11-14	88.50	93.03	85.91	88.51	85.97	88.15	88.88	88.60	88.78
22	DDCoT (T5)	PLM	Fine-tune	223M	223M	Link	23-10-25	88.72	86.84	84.91	87.59	83.34	88.08	88.58	85.10	87.34
23	LG-VQA (CLIP)	VLM	Zero-shot	-	-	Link	23-10-31	-	-	-	-	-	-	-	-	87.22
24	Chameleon (GPT-4)	Tool-LLM	Few-shot	1T+	-	Link	23-04-19	89.83	74.13	89.82	88.27	77.64	92.13	88.03	83.72	86.54
25	LG-VQA (BLIP-2)	VLM	Zero-shot	-	-	Link	23-10-31	-	-	-	-	-	-	-	-	86.32
26	LLaMA-SciTune	VLM	Zero-shot	7B	-	Link	23-07-03	84.50	94.15	82.91	88.35	83.64	88.74	85.05	85.60	86.11
27	Enigma-COT_Base	LLM	Fine-tune	229M	229M	Link	23-07-24	88.28	78.74	85.64	88.51	84.28	86.90	85.43	85.89	85.59
28	LLaMA-Adapter	VLM	Fine-tune	6B	1.2M	Link	23-03-28	84.37	88.30	84.36	83.72	80.32	86.90	85.83	84.05	85.19
29	Multimodal-CoT_Base	VLM	Fine-tune	223M	223M	Link	23-02-02	87.52	77.17	85.82	87.88	82.90	86.83	84.65	85.37	84.91
30	IMMO SL+RL	VLM	Fine-tune	7B	5M	Link	23-08-19	-	-	-	-	-	-	-	-	84.8
31	CoT GPT-4	LLM	Few-shot	1T+	-	Link	23-04-19	85.48	72.44	90.27	82.65	71.49	92.89	86.66	79.04	83.99
32	HoT-T5_Large	LLM	Fine-tune	738M	738M	Link	23-08-11	84.46	79.08	84.64	82.89	75.81	88.15	83.88	82.47	83.38
33	HoT-T5_Base	LLM	Fine-tune	223M	223M	Link	23-08-11	82.46	78.07	82.00	81.18	75.20	85.09	81.86	80.62	81.42
34	DDCoT (ChatGPT)	LLM	Zero-shot	175B	-	Link	23-10-25	80.15	76.72	82.82	78.89	72.53	85.02	82.86	75.21	80.15
35	Chameleon (ChatGPT)	Tool-LLM	Few-shot	175B+	-	Link	23-04-19	81.62	70.64	84.00	79.77	70.80	86.62	81.86	76.53	79.93
36	CoT GPT-3 + Doc	Tool-LLM	Zero-shot	173B	-	Link	23-08-01	-	-	-	-	-	-	-	-	79.91
37	UnifiedQA-T-SciQ_Base	LLM	Fine-tune	223M	223M	Link	23-05-05	76.56	88.99	80.45	72.90	73.84	83.47	81.09	75.19	79.41
38	CoT ChatGPT	LLM	Few-shot	175B	-	Link	23-04-19	78.82	70.98	83.18	77.37	67.92	86.13	80.72	74.03	78.31
39	DDCoT (GPT-3)	LLM	Zero-shot	175B	-	Link	23-10-25	78.60	73.90	80.45	77.27	69.96	82.93	80.65	73.50	78.09
40	LaVIN-13B	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	77.54
41	CoT GPT-3 (ALE)	LLM	Few-shot	173B	-	Link	22-09-20	75.44	70.87	78.09	74.68	67.43	79.93	78.23	69.68	75.17
42	LaVIN-7B	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	75.11
43	CoT GPT-3 (AE)	LLM	Few-shot	173B	-	Link	22-09-20	76.60	65.92	77.55	75.51	66.09	79.58	78.49	67.63	74.61
44	BLIP-2	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	74.17
45	CoT UnifiedQA	PLM	Fine-tune	223M	223M	Link	22-09-20	71.00	76.04	78.91	66.42	66.53	81.81	77.06	68.82	74.11
46	GPT-3 (0-shot)	LLM	Zero-shot	173B	-	Link	22-09-20	75.04	66.59	78.00	74.24	65.74	79.58	76.36	69.87	74.04
47	GPT-3 (2-shot)	LLM	Few-shot	173B	-	Link	22-09-20	74.64	69.74	76.00	74.44	67.28	77.42	76.80	68.89	73.97
48	InstructBLIP	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	73.33
49	UnifiedQA	PLM	Fine-tune	223M	223M	Link	22-09-20	68.16	69.18	74.91	63.78	61.38	77.84	72.98	65.00	70.12
50	ChatGPT	LLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	69.41
51	MetaCLIP	VLM	Zero-shot	-	-	Link	23-12-10	-	-	-	-	-	-	-	-	68.77
52	OpenCLIP	VLM	Zero-shot	-	-	Link	23-12-10	-	-	-	-	-	-	-	-	67.53
53	Flan-T5-XXL	LLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	67.43
54	SAM	VLM	Zero-shot	-	-	Link	23-12-10	-	-	-	-	-	-	-	-	67.08
55	DINOv2	VLM	Zero-shot	-	-	Link	23-12-10	-	-	-	-	-	-	-	-	64.60
56	VisualBERT	VLM	Fine-tune	111M	111M	Link	22-09-20	59.33	69.18	61.18	62.71	62.17	58.54	62.96	59.92	61.87
57	Patch-TRM	VLM	Fine-tune	90M	90M	Link	22-09-20	65.19	46.79	65.55	66.96	55.28	64.95	58.04	67.50	61.42
58	ViLT	VLM	Fine-tune	113M	113M	Link	22-09-20	60.48	63.89	60.27	63.20	61.38	57.00	60.72	61.90	61.14
59	DFAF	VQA-NN	Fine-tune	74M	74M	Link	22-09-20	64.03	48.82	63.55	65.88	54.49	64.11	57.12	67.17	60.72
60	Chat-UniVi	VLM	Zero-shot	7B	-	Link	23-11-14	58.61	61.08	61.82	57.33	58.25	61.39	62.04	56.23	59.96
61	BAN	VQA-NN	Fine-tune	112M	112M	Link	22-09-20	60.88	46.57	66.64	62.61	52.60	65.51	56.83	63.94	59.37
62	Top-Down	VQA-NN	Fine-tune	70M	70M	Link	22-09-20	59.50	54.33	61.82	62.90	54.88	59.79	57.27	62.16	59.02
63	MiniGPT4	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	58.70
64	LLaMA2-13B	LLM	Zero-shot	13B	-	Link	23-10-13	-	-	-	-	-	-	-	-	55.78
65	DDCoT (MiniGPT-4)	VLM	Zero-shot	-	-	Link	23-10-25	57.37	62.32	46.82	65.91	56.72	48.57	-	-	55.67
66	QVix	VLM	Zero-shot	-	-	Link	23-12-04	-	-	-	-	-	-	-	-	55.0
67	MCAN	VQA-NN	Fine-tune	95M	95M	Link	22-09-20	56.08	46.23	58.09	59.43	51.17	55.40	51.65	59.72	54.54
68	LLaMA-Adapter-V2	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	54.44
69	VLIS	VLM	Zero-shot	-	-	Link	23-10-15	-	-	-	53.1	49.3	49.1	-	-	50.2
70	LLaVA-13B	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	47.74
71	VPGTrans	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	47.00
72	MiniGPT-4	VLM	Zero-shot	-	-	Link	23-10-25	43.83	48.59	43.36	55.01	42.84	41.67	-	-	44.71
73	LLaMA1-13B	LLM	Zero-shot	13B	-	Link	23-10-13	-	-	-	-	-	-	-	-	43.33
74	LLaMA2-7B	LLM	Zero-shot	7B	-	Link	23-10-13	-	-	-	-	-	-	-	-	43.08
75	LLaVA-7B	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	41.10
76	OpenFlamingo	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	39.27
77	Lynx	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	38.28
78	mPLUG-Owl	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	37.93
79	MultiGPT	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	36.29
80	LLaMA1-7B	LLM	Zero-shot	7B	-	Link	23-10-13	-	-	-	-	-	-	-	-	36.19
81	Fromage	VLM	Zero-shot	-	-	Link	23-10-13	-	-	-	-	-	-	-	-	34.51
-	MMICL	VLM	Zero-shot	-	-	Link	23-09-14	-	-	-	-	82.10	-	-	-	-
-	VILA-13B (Llama-2-13B)	VLM	-	13B	-	Link	23-12-12	-	-	-	-	73.7	-	-	-	-
-	LLaVA+SIRI	VLM	-	13B	-	Link	23-11-29	-	-	-	-	72.0	-	-	-	-
-	LLaVA-1.5 (Vicuna-13B)	VLM	Zero-shot	13B	-	Link	23-10-05	-	-	-	-	71.6	-	-	-	-
-	InstructBLIP	VLM	Zero-shot	-	-	Link	23-09-14	-	-	-	-	71.30	-	-	-	-
-	ShareGPT4V (Vicuna-7B)	VLM	-	7B	-	Link	23-11-23	-	-	-	-	68.4	-	-	-	-
-	Qwen-VL-Chat (Qwen-7B)	VLM	Zero-shot	7B	-	Link	23-10-05	-	-	-	-	68.2	-	-	-	-
-	Multimodal Bard	VLM	Zero-shot	-	-	Link	23-08-07	-	-	-	-	68	-	-	-	-
-	Qwen-VL (Qwen-7B)	VLM	Zero-shot	7B	-	Link	23-10-05	-	-	-	-	67.1	-	-	-	-
-	LLaVA-1.5 (Vicuna-7B)	VLM	Zero-shot	7B	-	Link	23-10-05	-	-	-	-	66.8	-	-	-	-
-	Video-LLaVA (Vicuna-7B)	VLM	-	7B	-	Link	23-11-16	-	-	-	-	66.4	-	-	-	-
-	InstructBLIP+MoCLE	VLM	Zero-shot	7B	-	Link	23-12-19	-	-	-	-	63.9	-	-	-	-
-	Otter	VLM	Zero-shot	-	-	Link	23-09-14	-	-	-	-	63.10	-	-	-	-
-	InstructBLIP (Vicuna-13B)	VLM	Zero-shot	13B	-	Link	23-10-05	-	-	-	-	63.1	-	-	-	-
-	BLIP-2 (Vicuna-13B)	VLM	Zero-shot	13B	-	Link	23-10-05	-	-	-	-	61	-	-	-	-
-	InstructBLIP (Vicuna-7B)	VLM	Zero-shot	7B	-	Link	23-10-05	-	-	-	-	60.5	-	-	-	-
-	Ying-VLM	VLM	Zero-shot	-	-	Link	23-09-14	-	-	-	-	55.70	-	-	-	-
-	Shikra	VLM	Zero-shot	-	-	Link	23-09-14	-	-	-	-	45.80	-	-	-	-
	CaCo-CoT (Claude)	LLM	Few-shot	-	-	Link	23-08-23	-	-	-	90.8	-	-	-	-	-
	CaCo-CoT (Claude)	LLM	Zero-shot	-	-	Link	23-08-23	-	-	-	89.9	-	-	-	-	-
	CoT Claude	LLM	Few-shot	-	-	Link	23-08-23	-	-	-	89.5	-	-	-	-	-
	CaCo-CoT (ChatGPT)	LLM	Few-shot	-	-	Link	23-08-23	-	-	-	88.6	-	-	-	-	-
	Claude	LLM	Zero-shot	-	-	Link	23-08-23	-	-	-	86.8	-	-	-	-	-
	CaCo-CoT (ChatGPT)	LLM	Zero-shot	-	-	Link	23-08-23	-	-	-	86.5	-	-	-	-	-
	CoT Claude	LLM	Zero-shot	-	-	Link	23-08-23	-	-	-	86.5	-	-	-	-	-

Some notations in the table

• Model names:

  • Q: question
  • C: context
  • M: multiple options
  • A: answer
  • AE: answer with explanation
  • ALE: answer with lecture and explanation

• Method types:

  • VQA-NN: Standard neural network for visual question answering
  • PLM: Pre-trained language model
  • LLM: Large language model
  • VLM: (Large) vision-language model / Large multimodal model / Multimodal large language model
  • Tool-LLM: Tool-augmented large language model

• Learning:

  • Zero-shot: The model is evaluated in a zero-shot setting on ScienceQA
  • Few-shot: The model is evaluated in a few-shot setting on ScienceQA
  • Fine-tune: The model is fine-tuned on ScienceQA
  • -: Not available

• #Size: Total number of parameters in the model

• #P: Number of trainable parameters when fine-tuned on ScienceQA

• Accuracies for different question sets:

  • NAT: questions of the natural science subject
  • SOC: questions of the social science subject
  • LAN: questions of the language science subject
  • TXT: questions with the text context
  • IMG: questions with the image context
  • NO: questions with no context
  • G1-6: questions in the grade group of 1-6
  • G7-12: questions in the grade group of 7-12
  • Avg: all questions (reporting the average accuracy)

🗺️ About ScienceQA

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.

👻 Download the Dataset

The text part of the ScienceQA dataset is provided in data/scienceqa/problems.json.

The data examples in problem.json follow the following format:

{
  "(question id)": {
    "question": "question text, e.g., 'Which of these states is farthest north?'", 
    "choices": "question choices, e.g., ['West Virginia', 'Louisiana', 'Arizona', 'Oklahoma']",
    "answer": "index of ground truth answer, e.g., 0",
    "hint": "the textual context, e.g., 'Figure: sturgeon.', could be an empty string",
    "image": "the visual context with an image path, e.g., 'image.png', could be none",
    "task": "task type, e.g., 'closed choice'",
    "grade": "grade level, e.g., 'grade2'",
    "subject": "subject type, e.g., 'social science'",
    "topic": "topic type, e.g., 'geography'",
    "category": "category type, e.g., 'Geography'",
    "skill": "skill type, e.g., 'Read a map: cardinal directions'",
    "lecture": "lecture of the question, e.g., 'Maps have four cardinal directions, or main directions. Those directions are north, south, east, and west.\nA compass rose is a set of arrows that point to the cardinal directions. A compass rose usually shows only the first letter of each cardinal direction.\nThe north arrow points to the North Pole. On most maps, north is at the top of the map.'",
    "solution": "solution to the question, e.g., 'To find the answer, look at the compass rose. Look at which way the north arrow is pointing. West Virginia is farthest north.'",
    "split": "data split, e.g., 'train'"
  }
}

For language models such as UnifiedQA, GPT-3, and GPT-4, you can concatenate the textual hint and the captioning of the image to form the context.

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!

We have uploaded a demo to illustrate how to access the ScienceQA dataset on Hugging Face, available at demos/Hugging_Face_Dataset_demo.ipynb.

😈 Explore ScienceQA

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

🐙 Requirements

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

🤖 Run the GPT-3 (CoT) Model for ScienceQA

Generate the image captions

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

Run the model

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

Evaluate the results

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

Try different prompt templates

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

📖 Related Work

• 🔥 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]
• 🔥 TheoremQA: A Theorem-driven Question Answering Dataset, arXiv:2305.12524 [ArXiv] [Code] [Dataset]
• 🔥 SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models, arXiv:2307.10635 [ArXiv] [Code] [Dataset]
• 🔥 MathVista: Evaluating Math Reasoning in Visual Contexts with GPT-4V, Bard, and Other Large Multimodal Models, arXiv:2310.02255 [ArXiv] [Project] [Code] [Dataset]

⚠️ Licenses

This work is licensed under a MIT License.

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

✅ Cite

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}
}