lamm-mit/cephalo

Cephalo: Model Summary

Cephalo is a series of multimodal materials science focused vision large language models (V-LLMs) designed to integrate visual and linguistic data for advanced understanding and interaction in human-AI or multi-agent AI frameworks.

A novel aspect of Cephalo's development is the innovative dataset generation method. The extraction process employs advanced algorithms to accurately detect and separate images and their corresponding textual descriptions from complex PDF documents. It involves extracting images and captions from PDFs to create well-reasoned image-text pairs, utilizing large language models (LLMs) for natural language processing. These image-text pairs are then refined and validated through LLM-based NLP processing, ensuring high-quality and contextually relevant data for training.

Cephalo can interpret complex visual scenes and generating contextually accurate language descriptions and answer queries.

The models are developed to process diverse inputs, including images and text, facilitating a broad range of applications such as image captioning, visual question answering, and multimodal content generation. The architecture combines a vision encoder model and an autoregressive transformer to process complex natural language understanding.

Cephalo provides a robust framework for multimodal interaction and understanding, including the development of complex generative pipelines to create 2D and 3D renderings of material microstructures as input for additive manufacturing methods.

Overview of Models:

4b models

• Cephalo-Phi-3-vision-128k-4b-alpha
  • Base version of the Cephalo-Phi-3 model, trained on GPT-4o distilled image-text data from Wikipedia and scientific papers. Good baseline model, but struggles in longer conversations. Context length of 128,000 tokens.
• Cephalo-Phi-3-vision-128k-4b-beta
  • Improved version of the Cephalo-Phi-3 model, trained on GPT-4o and Idefics-2 distilled image-text data from Wikipedia and scientific papers, as well as a large text-only corpus. Provides nuanced responses, with excellent reasoning. Context length of 128,000 tokens.

8b models

• Cephalo-Idefics-2-vision-8b-alpha
  • Trained on Idefics-2 distilled image-text data from Wikipedia and scientific papers. Gives shorter answers, to the point, and generaly accurate.
• Cephalo-Idefics-2-vision-8b-beta
  • Trained on GPT-4o distilled image-text data from Wikipedia and scientific papers. Gives longer answers, with enhanced reasoning. Can struggle with complex concepts.
• Cephalo-Llava-v1.6-Mistral-8b-alpha
  • Trained on GPT-4o distilled image-text data from Wikipedia, with low-resolution images. Does not perform well on multiple image queries, and has some inconsistencies in understanding.

Merged 10b models

• Cephalo-Idefics-2-vision-10b-alpha
  • Merged model, 32+8=40 layers, checkpoint after first epoch. Trained on GPT-4o distilled image-text data from Wikipedia and scientific papers.
• Cephalo-Idefics-2-vision-10b-beta
  • Merged model, 32+8=40 layers, checkpoint after second epoch. Trained on GPT-4o distilled image-text data from Wikipedia and scientific papers.

Merged 12b models

• lamm-mit/Cephalo-Idefics-2-vision-12b-alpha
  • Merged model, 32+16=48 layers, checkpoint after first epoch. Trained on GPT-4o distilled image-text data from Wikipedia and scientific papers (dataset derivived from both Idefics-2 and GPT-4o distillation of the paper corpus).

The image shows a summary of model merging approach, constructing larger models from smaller pre-trained building blocks. a, Fine-tuning the base model. b, Constructing the larger, merged model by combining the whole or parts of smaller models. c, Fine-tuning the integrated hybrid, merged, model.

Mixture-of-Experts models

• lamm-mit/Cephalo-Phi-3-MoE-vision-128k-3x4b-beta

  • Mixture-of-expert model based on several smaller Cephalo-Phi-3 models. Provides a sample cookbook to make your own custom MoE vision models.

• lamm-mit/Cephalo-Idefics2-vision-3x8b-beta

  • Mixture-of-expert model based on several smaller Idefics-2 models. Provides a sample cookbook to make your own custom MoE vision models.

Etymology and inspiration behind the name Cephalo"

The name "Cephalo" is derived from the Greek word κεφαλή, or kephalē, meaning "head" or "brain", which symbolizes the model's central role in processing and integrating visual and linguistic information. This name reflects the model's function as the "brain" of the system, facilitating advanced human-AI and multi-agent AI interactions through the comprehensive understanding of multimodal data.

Additionally, "Cephalo" draws inspiration from cephalopods, a class of intelligent mollusks that includes octopuses, squids, and cuttlefish, associating it with the focus on biological inspiration that is central to the training and use of the model. Cephalopods are renowned for their exceptional cognitive abilities, advanced problem-solving skills, and highly developed nervous systems. They exhibit remarkable adaptability to their environments, sophisticated camouflage techniques, and complex behaviors, and are well-equipment to integrate visual cues with materialization.

By naming our multimodal materials science V-LLM "Cephalo", we evoke the intelligence and adaptability of cephalopods. Similar to how cephalopods process diverse sensory inputs to navigate and respond to their surroundings, Cephalo integrates and processes visual and linguistic data to handle complex tasks. This dual inspiration highlights the model's potential for advanced problem-solving and contextual comprehension, drawing parallels between the cognitive prowess of cephalopods and the model's capabilities in the realm of materials science and beyond.

Additional codes and tools

Additional codes and tools are provided at https://github.com/lamm-mit/Cephalo.

Citation

Please cite as:

@article{Buehler_Cephalo_2024,
  title={Cephalo: Multi-Modal Vision-Language Models for Bio-Inspired Materials Analysis and Design},
  author={Markus J. Buehler},
  journal={arxiv.org/abs/2405.19076},
  year={2024}
}