On November 12, 2026, NTT announced a technology that automatically generates diverse training data from a small amount of business documentation, enabling large language models (LLMs) to acquire specialized knowledge. In an evaluation using NTT's own business manuals, the accuracy of "tsuzumi 2" rose from 65.3% before training to 83.1%. The approach tackles a common problem: companies often cannot prepare enough documents for training. It uses an LLM to expand existing documents into more teaching material. However, needing fewer source documents does not automatically mean lower computing requirements or costs for training.
Automatically generating diverse training data from business manuals
NTT's new development is a tuning technique for adapting an existing LLM to specific business tasks. Tuning is also used to adjust a model's response format or its accuracy on particular tasks, but this technology aims to embed the specialized knowledge found in internal documents into the model.
The background is a large gap between the volume of business documents companies hold and the amount of data needed to train an LLM. According to NTT, retraining an entire model to give it new knowledge has traditionally involved data exceeding tens of gigabytes. Preparing that much data from the manuals for a single line of work is difficult. Collecting text from outside sources, meanwhile, risks mixing in information unrelated to the business. What was needed, then, was a way to increase training data while preserving the knowledge contained in limited internal documents.
NTT therefore developed a technology that has an LLM read business manuals used for tasks such as handling inquiries and automatically convert them into text in different formats, such as specifications and FAQs. Rather than repeatedly training the model on the original text, as in conventional methods, the LLM itself generates teaching material in multiple formats, and those materials are used to train the model.
For example, a specification describes functions and conditions systematically, while an FAQ organizes knowledge as pairs of questions and answers. Changing the format in which the same content is expressed makes it possible to create training data suited to different contexts and question styles. NTT says it generates the material with an eye to how knowledge is actually used in real work. A key feature is that companies can create varied training data from existing manuals without writing large volumes of new documents.
Learning from few documents is not the same as reducing computation
In evaluating this technology, it is important to distinguish between the amount of source documents a company prepares and the amount of data the model actually trains on.
In the evaluation graph (Figure 2) published by NTT, the vertical axis shows accuracy and the horizontal axis shows "training tokens (Billion)." A token is the basic unit an LLM uses to process text, and Billion means one billion. The horizontal axis is marked at 0.5, 5 and 50, comparing accuracy at different training data volumes. In other words, the graph shows the relationship between the number of tokens used for training and accuracy, not computation time or cost itself.
The process described in NTT's official announcement can be organized into three stages: preparing source documents, generating teaching material, and training the model. Each requires different data and resources.
| Stage | Target | What can be confirmed |
|---|---|---|
| Preparing source documents | Business manuals held by the company | A small amount of business documentation serves as the starting point for generating training data |
| Generating teaching material | Text in diverse formats such as specifications and FAQs | An LLM automatically generates training data from the source documents |
| Training the model | Number of tokens in the generated material | Training volume in units of billions of tokens is compared with accuracy |
This distinction matters because even when source documents are few, training the model on a large volume of material generated from them requires substantial computing resources. Reducing the documents a company must prepare is not the same as completing model training with less computation. The main advantage of this technology is that training data can be expanded from existing materials without collecting large amounts of source documents.
In addition, the number of training tokens alone cannot indicate GPU run time or cost. Assessing those requires information such as the model used to generate the material, the hardware configuration for training, and actual processing times. NTT's graph compares results for tsuzumi 2 and Llama 3 before training, with the new technology, and with prior research. To judge the relative merits of each method, one needs to check both the number of tokens used for training and the accuracy on the same evaluation questions.
tsuzumi 2 accuracy rises from 65.3% to 83.1%
In the results NTT announced, tsuzumi 2's accuracy on questions about the content of actual business manuals rose from 65.3% before training to 83.1%, an improvement of 17.8 points. The highest figure among the prior studies used for comparison was 73.3%, which the new technology exceeded by 9.8 points. That difference is our own calculation from the published figures; it shows improvement not only over the pre-training baseline but also over existing methods of generating teaching material.
The comparison methods include converting original documents into math problems or FAQs, as well as converting them into formats such as tables and procedure manuals. NTT also cites, as prior research, a method of rewriting text to suit reader personas randomly selected from sources such as Nemotron-Personas-Japan. The 73.3% figure is the highest accuracy among these prior studies. NTT says the new technology also outperformed prior research in experiments using Llama 3 (3B), and that it can be applied to LLMs other than tsuzumi.
However, the 83.1% accuracy was obtained on questions based on the business manuals NTT used. The announcement does not give details such as the number of questions, how the training and evaluation data were split, or statistical error. It is therefore impossible to say at this point whether similar improvements would be obtained with other companies' manuals or business content.
Furthermore, the number of pages and data size of the source manuals, the time required to generate the teaching material, and the prompts and data-selection criteria used in generation have not been disclosed. Implementation conditions for reproducing prior studies, as well as the GPU configuration, processing time and cost of training, have not been published either. The official announcement contains no link to a peer-reviewed paper on the method, and no independent replication has been found, so for now the results should be regarded as NTT's own R&D evaluation.
Companies considering adoption need to verify how accurately the tuned LLM answers the questions that actually arise in their own operations. The accuracy figure covers questions about what the manuals say; it does not directly show reductions in employees' working time or improvements in operational efficiency. If the conditions that allow the relationship between source documents, generated material and evaluation questions to be verified are made public, it will be easier for each company to assess whether it can expect similar gains.
Building on prior research, with a view to real-world deployment
The idea of using an LLM to convert existing text into different formats to increase training data is not itself new. "FinePhrase," published on arXiv by Joel Niklaus and colleagues, is research on generating synthetic data for pretraining by rephrasing web text. It compares text formats with various structures, including tables and math problems as well as FAQs and tutorials.
However, whereas FinePhrase deals with pretraining on general web text, NTT's technology aims to embed the specialized knowledge contained in corporate business manuals. Both the target text and the capability being evaluated differ. A distinguishing feature of this research is that it generates material with the use of knowledge in real work in mind and measures the effect using questions based on actual manuals. By comparing accuracy against multiple prior methods, it also shows concretely how much improvement is achieved when applied to business documents.
NTT positions the technology as one of several tuning methods to be combined, and plans to introduce it across NTT Group companies. It also plans to exhibit it at "NTT R&D FORUM 2026," to be held November 18–25, as part of efforts to broaden the uses of tsuzumi. The event will also introduce technology development in which AI agents proactively acquire business knowledge.
In practical use, the accuracy obtained with each company's own business documents will matter, as will the time and computing resources required to generate material and train the model. If the burden from data preparation to training can be kept down while securing the accuracy that operations demand, companies will have more options for developing LLMs suited to their own business without preparing large volumes of new specialized documents.