Intern-S1 - Shanghai AI Lab's open source scientific multimodal macromodels

What is Intern-S1?

Intern-S1 is a scientific multimodal grand model launched by Shanghai Artificial Intelligence Laboratory. The model deeply integrates linguistic and multimodal capabilities, and is equipped with powerful functions such as cross-modal scientific parsing, linguistic and visual fusion, scientific data processing, scientific question answering, experiment design and optimization, and so on. Intern-S1 is the first "cross-modal scientific parsing engine", which can accurately interpret complex scientific modal data such as chemical molecular formulas, protein sequences, seismic signals, etc., and outperforms the top closed-source models in multidisciplinary professional tasks. Intern-S1 is based on Dynamic Tokenizer and Time Series Signal Encoder to realize the deep fusion of multiple scientific modalities, and the scientific data synthesis method of generalization and specialization fusion is equipped with powerful general reasoning ability and top professional ability, which is widely used in the scenarios of scientific research tool integration, image and text fusion, and complex scientific modal data processing.

Main functions of Intern-S1

• Cross-modal scientific analysisThe following are some examples: accurate interpretation of complex scientific modal data, such as predicting compound synthesis pathways and determining the feasibility of chemical reactions in chemistry; assisting in analyzing protein sequences and discovering drug targets in biomedicine; and identifying seismic wave signals and analyzing seismic events in the field of earth sciences.
• Verbal and visual integration: Combining verbal and visual information for complex multimodal tasks.
• Scientific data processing: It supports a variety of complex scientific modal data inputs, such as light-change curves in materials science, gravitational wave signals in astronomy, etc., and realizes in-depth fusion and efficient processing of data.
• Answers to scientific questions: Provide users with accurate answers to scientific questions based on a powerful knowledge base and reasoning capabilities.
• Experimental design and optimization: Assist researchers in designing experimental protocols, optimizing experimental processes, and improving research efficiency.
• multi-intelligence unit (MIM) synergy: Supports multi-intelligent body systems that work in concert with other intelligences to accomplish complex scientific tasks.
• Autonomous Learning and Evolution: It has the ability to learn on its own and optimize its performance based on its interaction with the environment.
• Data processing and analysis: Provides data processing and analysis tools to help researchers quickly process and analyze scientific data.
• Model deployment and application: Supports local deployment and cloud services for easy use in different scenarios.

Intern-S1 official website address

• Project website:: https://intern-ai.org.cn/
• Github repository:: https://github.com/InternLM/Intern-S1
• HuggingFace Model Library:: https://huggingface.co/internlm/Intern-S1-FP8

How to use Intern-S1

• Online Experience Platform
  • Access platforms: Open your browser and visit the project's official website to experience the Intern-S1 model.
  • Starting a conversation: Enter a question or request in the input box and submit to get a response from Intern-S1.
  • Use of special features: Follow the instructions on the platform according to the function of interest, e.g. organic chemistry.
• GitHub repository
  • clone warehouse: Clone the repository by typing the following command at the command line:

git clone https://github.com/InternLM/Intern-S1.git

• • Installation of dependencies: Go to the repository directory and install the Python dependencies:

cd Intern-S1
pip install -r requirements.txt

• • operational model: Run the model based on the README file in the repository or the sample code. Typically, you can use Python scripts:

python script_name.py

• Hugging Face Model Library
  • Loading Models: Use the following code to load the model and the disambiguator in the Python environment:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "internlm/Intern-S1-FP8"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

• • Generate Text: Use the model to generate text or perform other tasks:

inputs = tokenizer("Tell me about an interesting physical phenomenon.", return_tensors="pt")
output = model.generate(**inputs, max_length=100)
print(tokenizer.decode(output[0], skip_special_tokens=True))

• • Using the Model API: If Hugging Face provides an API service, call the model directly through the API:

import requests

url = "https://api-inference.huggingface.co/models/internlm/Intern-S1-FP8"
headers = {"Authorization": "Bearer YOUR_HUGGINGFACE_API_TOKEN"}
data = {"inputs": "Tell me about an interesting physical phenomenon."}

response = requests.post(url, headers=headers, json=data)
print(response.json()["generated_text"])

Core Benefits of Intern-S1

• multimodal capability: Combining powerful language understanding and visual processing capabilities, it is able to process and understand text, images and many other types of data.
• Scientific mission optimization: Models excel at tasks in the sciences, such as deciphering chemical structures, understanding protein sequences, and planning compound synthesis pathways.
• Large-scale pre-training: Built on a 235 billion-parameter MoE (Mixture of Experts) language model and a 6 billion-parameter visual coder, pre-trained on 5 trillion tokens of multimodal data, of which more than 2.5 trillion tokens are from the scientific domain.
• Dynamic Tokenizer: The model is based on a dynamic Tokenizer that supports native understanding of specialized data such as molecular formulas, protein sequences, and seismic signals.
• Cross-modal scientific parsing engine: Intern-S1's pioneering cross-modal scientific parsing engine can accurately interpret a wide range of complex scientific modal data, demonstrating excellent scientific reasoning and comprehension capabilities.
• General Reasoning and Professional Competence: Based on the scientific data synthesis method of generalization and specialization fusion, Intern-S1 has powerful general reasoning ability and possesses a number of top professional capabilities.
• Autonomous learning and evolution: The model has the ability to learn on its own, and can continuously optimize its performance by interacting with the environment.
• multi-intelligence unit (MIM) synergy: Supports multi-intelligent body systems that can work in concert with other intelligences to accomplish complex scientific tasks.
• Data processing and analysis: Provides data processing and analysis tools to help researchers quickly process and analyze scientific data.

People for whom Intern-S1 is intended

• (scientific) researcher: Researchers who perform complex data analysis and experimental design in the fields of chemistry, biology, physics, and earth sciences.
• Data Analyst: Professionals who extract valuable information from large amounts of scientific data to support decision making.
• software developer: Software engineers who integrate advanced multimodal analysis capabilities into their applications.
• Educators and students: Teachers and students who use models to assist in teaching or conducting scientific research in an academic setting.
• Corporate decision makers: Business leaders who make critical business decisions based on data analytics.