Starting custom service of SLM fine-tuning to improve the answer accuracy of AI generated using small language models such as Phi-3, Llama-3, and GPT-4o mini.
Headwaters Co., Ltd. (Headquarters: Shinjuku-ku, Tokyo, Representative Director: Yosuke Shinoda, hereinafter referred to as Headwaters), which handles AI solution business, has started offering the custom service of SLM fine-tuning for companies that promote the use of generated AI for business operations.
This service uses small language models centered on open source AI platform models such as Phi-3, Llama-3, and GPT-4o mini that can be selected from the Azure AI Model Catalog provided by Microsoft Corporation to improve the answer accuracy of generated AI. This service is useful for companies that are considered difficult to use in business due to the accuracy of the sentences created by generated AI.
Headwaters has developed many solutions, such as generated AI and LLM (Large Language Model), RAG (Retrieval Augmented Generation) system using Headwaters' technical capabilities, and edge AI using SLM (Small Language Model). We have expanded the lineup of GPT services for companies using the Azure OpenAI Service, and have been exploring solutions to common issues such as 'wanting to support specialized terms, industry terms, and in-house terms,' 'wanting to suggest and recommend when specific keywords appear,' and 'wanting to improve answer accuracy' from many customers who use generated AI for business operations.
One common challenge in the business use of generated AI is 'wanting to support specialized terms, industry terms, and in-house terms,' 'wanting to suggest and recommend when specific keywords appear,' and 'wanting to improve answer accuracy.' Headwaters has been exploring solutions to these issues.
In response to such requests, Headwaters has launched a custom service of SLM fine-tuning centered on 'Phi-3' of Microsoft 'SLM,' 'Llama-3' of Meta, and 'GPT-4o mini' of OpenAI.
Issues with RAG
Usually, to use LLM for business use, each company needs to customize LLM based on its business data or unique prompts. Customization methods are divided into two types: RAG and fine-tuning. However, because fine-tuning requires a high level of difficulty and effective data, and is costly, it is common to start with RAG, which has a well-balanced cost performance.
On the other hand, some of the issues with RAG are 'hallucination (incorrect answer) due to referring to too much data,' 'response to cases where it is desired to increase the accuracy of generated AI,' 'corresponding to in-house terms, industry terms, and specialized terms that are not well known to the general public.'
To address these issues, Headwaters will use SLM to solve the problem by utilizing SLM for 'in-house terms, industry terms, and specialized terms' and 'increasing the accuracy of the answers.'
Features of SLM
The main feature of SLM is 'reducing the weight of LLM,' but another feature is the 'small amount of data handled.'
By preparing 'industry-specific terms and nuances' and 'answers that should be given priority over other knowledge, such as answers that must not be incorrect' as SLM learning data, the risk of generating inaccuracies and irrelevant information can be minimized. Furthermore, compared to LLM, SLM can save on computing resources, which makes it a cost-effective and efficient solution that can reduce response time and energy consumption.
In addition, to improve the affinity with Microsoft Azure and Copilot+ PC, Headwaters uses 'Phi-3' provided by Microsoft to consider the affinity with Microsoft Azure and Copilot+ PC, and uses a Japanese-learning model developed based on 'Llama-3' of Meta, which solves the weakness of SLM, which is considered to be the Japanese correspondence. To solve the issues in terms of operating costs and speed, we use 'GPT-4o mini' of OpenAI.
By using SLM, which handles a small amount of data, it is now possible to provide fine-tuning, which would otherwise be expensive, at a lower cost than LLM fine-tuning.
Fine tuning requires expertise in data science, but by combining machine learning knowledge accumulated over the years and expertise of Kaggle medalists with implicit and industry-specific terminology separation method, there are several introduction cases and improvement of accuracy has been confirmed.
At Headwaters, we strive for further cost performance improvement by providing Advanced RAG services using SLM fine tuning and Microsoft Fabric as well as a generative AI platform "SyncLect Generative AI" composed of Microsoft Azure and SLM fine tuning, and business utilization of generative AI in enterprise companies such as manufacturing, finance, broadcasting, health care and support for customer service platforms utilizing generative AI with relatively high correct answer rates.
■ Future prospects
In the future, we will expand the SLM service lineup and realize the following solution deployments.
- Multimodal SLM "GPT-4o mini," "Phi-3 Vision," and "Florence-2" for multitask edge video analysis
- Generating AI x on-premises that does not take out personal information to the cloud
- Local SLM that supports offline environments
- Copilot+ Windows AI application that runs on a PC
- On-device SLM application that runs on mobile devices ...etc
Headwaters regards alliance strategy as one of the pillars of medium- to long-term strategy, and is working to expand the generative AI economy in partnership with customer companies. We will incorporate generative AI into customer business and exchange customers to bring generative AI closer to a world where it is used naturally and close.
Please use your Futubull account to access the feature.
SLM (Small Language Model) is a language model that is smaller in size and lighter weight than LLM (Large Language Model). It enables high-speed training and inference, enhances resource efficiency, and excels in cost performance. In addition, there are various features such as being suitable for devices with limited resources and edge computing, and being secure and highly confidential. The potential for smaller language models is attracting attention in the generative AI category, and the adoption of small language models is increasing.
SLM (Small Language Model) is a language model that is smaller in size and lighter weight than LLM (Large Language Model). It enables high-speed training and inference, enhances resource efficiency, and excels in cost performance. In addition, there are various features such as being suitable for devices with limited resources and edge computing, and being secure and highly confidential. The potential for smaller language models is attracting attention in the generative AI category, and the adoption of small language models is increasing.
Fine-tuning is a method of adding new layers to an already trained model and retraining the entire model. By reusing the model, it is possible to construct the model with less data in a shorter time than learning from scratch.
Fine-tuning is a method of adding new layers to an already trained model and retraining the entire model. By reusing the model, it is possible to construct the model with less data in a shorter time than learning from scratch.
Phi-3 is an open source small language model (SLM) provided by Microsoft. It demonstrates the highest level of ability and cost efficiency, exceeding models of equal and the next size in various languages, inference, coding, and mathematical benchmarks.
GPT-4o mini is a small model of the multimodal language model "GPT-4o" provided by to OpenAI. It is a model that is more than 60% cheaper for developers to use than GPT3.5 and has significantly increased its speed in addition to improving its accuracy.
Azure AI Model Catalog is a packaged top-level infrastructure model that accelerates development of major open-source generative AI models such as OpenAI, Meta, Mistral AI, Stability AI, and Hugging Face, etc.
GPT-4o mini is a small model of the multimodal language model "GPT-4o" provided by to OpenAI. It is a model that is more than 60% cheaper for developers to use than GPT3.5 and has significantly increased its speed in addition to improving its accuracy.
Azure AI Model Catalog is a packaged top-level infrastructure model that accelerates development of major open-source generative AI models such as OpenAI, Meta, Mistral AI, Stability AI, and Hugging Face, etc.
Azure AI Model Catalog is a packaged top-level infrastructure model that accelerates development of major open-source generative AI models such as OpenAI, Meta, Mistral AI, Stability AI, and Hugging Face, etc.
RAG (Retrieval Augmented Generation) is ...
Retrieval Augmented Generation (RAG) is a new technology that combines large language models (LLMs) with external databases and information sources. It searches external knowledge sources for more enhanced article generation.
What is Copilot+ PC?
Copilot+ PC is a new class of Windows 11 PC designed specifically for process-intensive tasks such as real-time translation and image generation, utilizing AI that features a super high-speed neural processing unit (NPU) capable of executing over 40 trillion operations per second (TOPS) and equipped with advanced computer chips.
Reference
Microsoft Fabric-based Advanced RAG service launched.
o.jp/news/gen_ai_microsoft_fabric_advanced_rag.html
Development of LLaVA Edge Vision, an industrial edge generation AI solution.
Verification of small language models (SLMs) and image language models (VLMs) for generated AI x edge AI.
Advanced partner certification for Azure OpenAI Service reference architecture.
■ Trademarks
Microsoft, Windows and Azure are registered trademarks or trademarks of Microsoft Corporation in the United States and other countries.
The official name of Windows is Microsoft Windows Operating System.
Other proper nouns such as product names mentioned are trademarks or registered trademarks of their respective companies.
Founded: November 2005
URL:
Company Name: Headwaters Co., Ltd. Location: Shinjuku Island Tower 4th Floor, 6-5-1 Nishishinjuku, Shinjuku-ku, Tokyo 163-1304 Representative Director: Yusuke Shinoda Established: November 2005 URL: https://www.headwaters.co.jp
Address: Shinjuku Island Tower 4th Floor, 6-5-1 Nishishinjuku, Shinjuku-ku, Tokyo 163-1304
Representative Director: Yusuke Shinoda
Headwaters Co., Ltd.
Email: info@ml.headwaters.co.jp
This service uses small language models centered on open source AI platform models such as Phi-3, Llama-3, and GPT-4o mini that can be selected from the Azure AI Model Catalog provided by Microsoft Corporation to improve the answer accuracy of generated AI. This service is useful for companies that are considered difficult to use in business due to the accuracy of the sentences created by generated AI.
Headwaters has developed many solutions, such as generated AI and LLM (Large Language Model), RAG (Retrieval Augmented Generation) system using Headwaters' technical capabilities, and edge AI using SLM (Small Language Model). We have expanded the lineup of GPT services for companies using the Azure OpenAI Service, and have been exploring solutions to common issues such as 'wanting to support specialized terms, industry terms, and in-house terms,' 'wanting to suggest and recommend when specific keywords appear,' and 'wanting to improve answer accuracy' from many customers who use generated AI for business operations.
One common challenge in the business use of generated AI is 'wanting to support specialized terms, industry terms, and in-house terms,' 'wanting to suggest and recommend when specific keywords appear,' and 'wanting to improve answer accuracy.' Headwaters has been exploring solutions to these issues.
In response to such requests, Headwaters has launched a custom service of SLM fine-tuning centered on 'Phi-3' of Microsoft 'SLM,' 'Llama-3' of Meta, and 'GPT-4o mini' of OpenAI.
Issues with RAG
Usually, to use LLM for business use, each company needs to customize LLM based on its business data or unique prompts. Customization methods are divided into two types: RAG and fine-tuning. However, because fine-tuning requires a high level of difficulty and effective data, and is costly, it is common to start with RAG, which has a well-balanced cost performance.
On the other hand, some of the issues with RAG are 'hallucination (incorrect answer) due to referring to too much data,' 'response to cases where it is desired to increase the accuracy of generated AI,' 'corresponding to in-house terms, industry terms, and specialized terms that are not well known to the general public.'
To address these issues, Headwaters will use SLM to solve the problem by utilizing SLM for 'in-house terms, industry terms, and specialized terms' and 'increasing the accuracy of the answers.'
Features of SLM
The main feature of SLM is 'reducing the weight of LLM,' but another feature is the 'small amount of data handled.'
By preparing 'industry-specific terms and nuances' and 'answers that should be given priority over other knowledge, such as answers that must not be incorrect' as SLM learning data, the risk of generating inaccuracies and irrelevant information can be minimized. Furthermore, compared to LLM, SLM can save on computing resources, which makes it a cost-effective and efficient solution that can reduce response time and energy consumption.
In addition, to improve the affinity with Microsoft Azure and Copilot+ PC, Headwaters uses 'Phi-3' provided by Microsoft to consider the affinity with Microsoft Azure and Copilot+ PC, and uses a Japanese-learning model developed based on 'Llama-3' of Meta, which solves the weakness of SLM, which is considered to be the Japanese correspondence. To solve the issues in terms of operating costs and speed, we use 'GPT-4o mini' of OpenAI.
By using SLM, which handles a small amount of data, it is now possible to provide fine-tuning, which would otherwise be expensive, at a lower cost than LLM fine-tuning.
Fine tuning requires expertise in data science, but by combining machine learning knowledge accumulated over the years and expertise of Kaggle medalists with implicit and industry-specific terminology separation method, there are several introduction cases and improvement of accuracy has been confirmed.
At Headwaters, we strive for further cost performance improvement by providing Advanced RAG services using SLM fine tuning and Microsoft Fabric as well as a generative AI platform "SyncLect Generative AI" composed of Microsoft Azure and SLM fine tuning, and business utilization of generative AI in enterprise companies such as manufacturing, finance, broadcasting, health care and support for customer service platforms utilizing generative AI with relatively high correct answer rates.
■ Future prospects
In the future, we will expand the SLM service lineup and realize the following solution deployments.
- Multimodal SLM "GPT-4o mini," "Phi-3 Vision," and "Florence-2" for multitask edge video analysis
- Generating AI x on-premises that does not take out personal information to the cloud
- Local SLM that supports offline environments
- Copilot+ Windows AI application that runs on a PC
- On-device SLM application that runs on mobile devices ...etc
Headwaters regards alliance strategy as one of the pillars of medium- to long-term strategy, and is working to expand the generative AI economy in partnership with customer companies. We will incorporate generative AI into customer business and exchange customers to bring generative AI closer to a world where it is used naturally and close.
Please use your Futubull account to access the feature.
SLM (Small Language Model) is a language model that is smaller in size and lighter weight than LLM (Large Language Model). It enables high-speed training and inference, enhances resource efficiency, and excels in cost performance. In addition, there are various features such as being suitable for devices with limited resources and edge computing, and being secure and highly confidential. The potential for smaller language models is attracting attention in the generative AI category, and the adoption of small language models is increasing.
SLM (Small Language Model) is a language model that is smaller in size and lighter weight than LLM (Large Language Model). It enables high-speed training and inference, enhances resource efficiency, and excels in cost performance. In addition, there are various features such as being suitable for devices with limited resources and edge computing, and being secure and highly confidential. The potential for smaller language models is attracting attention in the generative AI category, and the adoption of small language models is increasing.
Fine-tuning is a method of adding new layers to an already trained model and retraining the entire model. By reusing the model, it is possible to construct the model with less data in a shorter time than learning from scratch.
Fine-tuning is a method of adding new layers to an already trained model and retraining the entire model. By reusing the model, it is possible to construct the model with less data in a shorter time than learning from scratch.
Phi-3 is an open source small language model (SLM) provided by Microsoft. It demonstrates the highest level of ability and cost efficiency, exceeding models of equal and the next size in various languages, inference, coding, and mathematical benchmarks.
GPT-4o mini is a small model of the multimodal language model "GPT-4o" provided by to OpenAI. It is a model that is more than 60% cheaper for developers to use than GPT3.5 and has significantly increased its speed in addition to improving its accuracy.
Azure AI Model Catalog is a packaged top-level infrastructure model that accelerates development of major open-source generative AI models such as OpenAI, Meta, Mistral AI, Stability AI, and Hugging Face, etc.
GPT-4o mini is a small model of the multimodal language model "GPT-4o" provided by to OpenAI. It is a model that is more than 60% cheaper for developers to use than GPT3.5 and has significantly increased its speed in addition to improving its accuracy.
Azure AI Model Catalog is a packaged top-level infrastructure model that accelerates development of major open-source generative AI models such as OpenAI, Meta, Mistral AI, Stability AI, and Hugging Face, etc.
Azure AI Model Catalog is a packaged top-level infrastructure model that accelerates development of major open-source generative AI models such as OpenAI, Meta, Mistral AI, Stability AI, and Hugging Face, etc.
RAG (Retrieval Augmented Generation) is ...
Retrieval Augmented Generation (RAG) is a new technology that combines large language models (LLMs) with external databases and information sources. It searches external knowledge sources for more enhanced article generation.
What is Copilot+ PC?
Copilot+ PC is a new class of Windows 11 PC designed specifically for process-intensive tasks such as real-time translation and image generation, utilizing AI that features a super high-speed neural processing unit (NPU) capable of executing over 40 trillion operations per second (TOPS) and equipped with advanced computer chips.
Reference
Microsoft Fabric-based Advanced RAG service launched.
o.jp/news/gen_ai_microsoft_fabric_advanced_rag.html
Development of LLaVA Edge Vision, an industrial edge generation AI solution.
Verification of small language models (SLMs) and image language models (VLMs) for generated AI x edge AI.
Advanced partner certification for Azure OpenAI Service reference architecture.
■ Trademarks
Microsoft, Windows and Azure are registered trademarks or trademarks of Microsoft Corporation in the United States and other countries.
The official name of Windows is Microsoft Windows Operating System.
Other proper nouns such as product names mentioned are trademarks or registered trademarks of their respective companies.
Company Name: Headwaters Co., Ltd. Location: Shinjuku Island Tower 4th Floor, 6-5-1 Nishishinjuku, Shinjuku-ku, Tokyo 163-1304 Representative Director: Yusuke Shinoda Established: November 2005 URL: https://www.headwaters.co.jp
Address: Shinjuku Island Tower 4th Floor, 6-5-1 Nishishinjuku, Shinjuku-ku, Tokyo 163-1304
Representative Director: Yusuke Shinoda
Headwaters Co., Ltd.
Email: info@ml.headwaters.co.jp
Previous article