SPHINX-X: Scaling Data and Parameters for a Family of Multi-modal Large Language Models

By Peng Gao, Renrui Zhang, Chris Liu, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin and 9 more. 2024

159 citations recorded by the source corpus.

We propose SPHINX-X, an extensive Multimodality Large Language Model (MLLM) series developed upon SPHINX. To improve the architecture and training efficiency, we modify the SPHINX framework by removing redundant visual encoders, bypassing fully-padded sub-images with skip tokens, and simplifying multi-stage training into a one-stage all-in-one paradigm. To fully unleash the potential of MLLMs, we assemble a comprehensive multi-domain and multimodal dataset covering publicly available resources in language, vision, and vision-language tasks. We further enrich this collection with our curated OCR intensive and Set-of-Mark datasets, extending the diversity and generality. By training over different base LLMs including TinyLlama1.1B, InternLM2-7B, LLaMA2-13B, and Mixtral8x7B, we obtain a spectrum of MLLMs that vary in parameter size and multilingual capabilities. Comprehensive benchmarking reveals a strong correlation between the multi-modal performance with the data and parameter scales. Code and models are released at https://github.com/Alpha-VLLM/LLaMA2-Accessory

Subfields: Artificial Intelligence, Large Language Models, Multimodal Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Hongsheng Li
2. Yu Qiao
3. Peng Gao
4. Saining Xie
5. Bo Li
6. Dahua Lin
7. Yoshua Bengio
8. Dacheng Tao

In Large Language Models

1. Bo Li
2. Yu Qiao
3. Peng Gao
4. Jiaming Han
5. Yann LeCun
6. Dahua Lin
7. Saining Xie
8. Hao Shao

In Multimodal Learning

1. Hongsheng Li
2. Peng Gao
3. Yu Qiao
4. Renrui Zhang
5. Pan Lu
6. Saining Xie
7. Penghao Wu
8. Yuhang Zang

Citations in the corpus

• LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention (2023, background): “The current working directory is /home/user/ echo “I am a chatbot“ > file.txt && cat file.txt && ls -l I am a chatbot -rw-r–r– 1 user user”
• LMDrive: Closed-Loop End-to-End Driving with Large Language Models (2023, background): “Furthermore, vision large language models (VLLM) [15, 28, 40] further introduce vision encoders and open the doors for LLMs to interpret not only textual data but also images and data in other modalities [5, 29, 58].”
• Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents (2024, unknown): “LLaVA (Liu et al., 2024c;a) was initially built for 336 px images, and was later scaled up to at most 772 px via the AnyRes technique (Cheng et al., 2023; Gao et al., 2024; Liu et al., 2024b; Guo et al., 2024; Dong et al., 2024).”
• MathVerse: Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems? (2024, methodology): “5 [37] (Vicuna-13B [13]), LLaVA-NeXT [38] (Vicuna-13B), SPHINX-MoE [21] (Mixtral-8×7B [28]), SPHINX-Plus (LLaMA2-13B [56]), InternLM-XComposer2 [16] (InternLM2-7B [54]), LLaMA-Adapter V2 [20] (LLaMA-7B [55]), ImageBind-LLM [24] (LLaMA-7B), MiniGPT-v2 [11] (LLaMA2-7B), mPLUG-Owl2 [62] (LLaMA-7B),…”
• LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models (2024, background): “Recent advancements in Large Multimodal Models (LMMs) [11, 12, 26, 37, 43, 64] have showcased impressive capabilities in diverse multimodal contexts, advancing the pursuit of artificial general intelligence.”
• Are We on the Right Way for Evaluating Large Vision-Language Models? (2024, background): “2 [6], Sphinx-X-8x7B [15].”
• Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs (2024, background): “There are also mixed practices in freezing [56, 73, 75] or unfreezing [38, 74] vision backbones during fine-tuning.”
• Visual CoT: Advancing Multi-Modal Language Models with a Comprehensive Dataset and Benchmark for Chain-of-Thought Reasoning (2024, methodology): “This enthusiasm has led to the emergence of multi-modal large language models (MLLM), such as LLaVA [33, 34], SPHINX [14, 31], and Qwen-VL [3].”