BAOBAB

Comprehensive LLM Services Including Everything from Training Data Creation to Model Evaluation

We provide rapid delivery of high-quality tuning data sets for building large language models (LLMs) that specialise in specific tasks and applications. We will support you along every step of the way, starting from the requirement establishment stage.

Your Reliable Partner for LLM Development

  • Over 10 years' experience

    Building on over a decade’s worth of knowledge and expertise in building textual data sets, we are able to assemble teams specialised in performing specific tasks, and carry out efficient operations. This allows us to rapidly deliver high-quality data sets that suit your individual requirements.

  • Support right from the initial requirements stage

    Thoroughly establishing specific requirements is essential for the creation of high-quality data. We offer support and advice in establishing specifications for training data that are in line with the purpose of the model.

  • Ethical AI

    We have established a code of ethics and wage regulations for our partners (Baoparts), ensuring appropriate working conditions and upholding human rights for the people who carry out work for us, resulting in ethical data creation. We are contributing to building a fair and healthy AI ecosystem.

Baobab's LLM Services

Data set creation for fine tuning

  • Fine tuning is essential for building large language models (LLMs) that specialise in specific tasks and applications. At Baobab, we discuss with clients in detail regarding the purpose and application of the model to be fine-tuned, and use this information to create high-quality fine-tuning data sets that are consistent and accurate.

Data set creation for retrieval-augmented generation (RAG)

  • To use RAG in an LLM, you need to not only engineer the right kind of prompts to implement RAG, you must also have a high-quality data set to tune the LLM for RAG. Such a data set will typically contain the following information:

    • - The text of users' questions
    • - Queries to extract the information matching the users' questions from the information source
    • - The information extracted from the information source
    • - The language model's answers

      • Building on over a decade’s worth of knowledge and expertise in building textual datasets, we are able to rapidly deliver high-quality data sets that suit clients' individual requirements. We are also able to create RAG data sets for private data by contractual agreement. We can create and deliver data sets for customers to tune their own models, and evaluate models that are in operation.

Download sample RAG data set (free)

Reinforcement learning from human feedback (RLHF)

  • The RLHF process requires a large amount of question data and question and answer data sets created manually by humans. In addition, ranking the responses extracted from the tuned model must also be performed by humans. At Baobab, we create a QA data set that suits the model the client is aiming for, and rank the sampled results.

Baobab generative AI evaluation service (manual evaluation service)

  • We provide manual evaluation of the output from LLMs, etc., undertaken by Baobab evaluators. This can help to avoid problems such as prejudices and observer bias, and ensure the development of trustworthy generative AI with guaranteed objectivity.

Baobab's LLM Service Accomplishments

  • Internet service

    Instruction/response annotation

    Issue
    The client was aiming to build an LLM that understands Japanese and the sensitivity of Japanese people, and was seeking the speedy creation of training data.
    Service provided
    After discussing and consulting with the client, our experts proposed a training data creation strategy that was aligned with their business objective. We provided instruction response text totalling 390,000 characters in 12 business days. During the project, any questions were shared with the client on a question sheet, and efforts were made to ensure that there were no misunderstandings.
    Outcome
    The client was satisfied with the rapid response and high-quality data, and said they had good results using it to train their model. Consequently, the client is also considering Baobab for additional data and data construction for reinforcement learning in the future.
  • Global manufacturer

    Data set creation for evaluation

    Issue
    In order to build an LLM dialogue model specialised in Q&A regarding the manufacturer's technical manual for engineers, a large amount of manually created data was required for training and evaluation.
    Service provided
    Based on interviews with the client, we assembled a team led by an advisor with the specific technical expertise required for the data, and established a system that enabled us to check whether technical terminology was correct. Furthermore, devising a communication method with the client during the preliminary stages before work started enabled us to have a firmer understanding of the information, which enhanced the quality of the deliverables.
    Outcome
    The client was satisfied with the extremely high quality of the data set, which was achieved thanks to team members with specialist knowledge participating in the project, and because the justification for the data had been made clear.

Project process

In the case when a contract is established directly between Baobab and the client requesting creation of a data set

1) Project consultation

We hold consultation to gather details of the request and assess whether we can take on the project.

2) Conclusion of NDA

We enter into a non-disclosure agreement with the client.

3) Establish data set specifications

We establish specifications for the data set together with the client.

4) Provision of information

The client provides us with the information source required to construct the data set.

5) Data creation

We inform our annotators about the task and train them for it, then create the data set.

6) Delivery

We inspect and deliver the completed data set.