TL;DR: We automate generating training & evaluation data in technical domains, primarily in the format of prompt/response pairs. Our API is in closed beta, but you can play with a stripped-down demo here!

‍

A team we work with was using financial analysts to write prompt/answer pairs to improve their LLM agents. They'd write questions like “How does Dodd-Frank legislation affect X company action”, or “What filing would you need to make if an entity did Y?”. This process, however, took hours and was a drag for employees.

But what if we didn’t actually need humans to create labeled data?

Sometimes we’re not looking for net new knowledge. For example when we were trying to encode financial regulation into a model, all the knowledge we need is already written up. In this case, generating a dataset is not a problem of finding new knowledge, but transforming existing knowledge. 

The challenge then becomes, “How do we transform unstructured knowledge into rows of data?”

Similar methods have been adopted by teams like Microsoft with AdaptLLM, where they turn sentences into reading comprehension tasks, and then use the resulting tasks as training data for a general purpose LLM. We expand on these types of approaches by parsing unstructured text to form knowledge graphs, and turning those knowledge graphs into datasets.

At a high level, we:

1. Parse existing documents for interesting facts and lessons we want to know.
2. Reverse engineer questions or situations that require those facts to answer properly
3. Transform the question/answer to closely match as much production data as we can in tone, style, and format.

Example

For example, here’s a few snippets of banking regulation [1]

1. “G-SIBs, or Global Systemically Important Banks, is a list of institutions identified annually by the Financial Stability Board (FSB)... that includes JP Morgan, Wells Fargo…..”
2. “G-SIBs are required to hold a minimum of 4.5% CET1 holdings”
3. “Common Equity Tier 1 (CET1) is the sum of common shares and stock surplus”

From these snippets, we could parse this fact:‍

       JP Morgan and Well Fargo must hold at least 4.5% of its assets in common stock and cash

Now we can generate multiple scenarios that demand advanced financial knowledge:

Question: JP Morgan currently holds 5% of its risk weighted assets in cash and common stock; is this ratio compliant with federal regulations?

‍Answer: Since JPM is a GSIB it is required to hold 4.5% of assets in cash and stock. Therefore this holding is complaint with regulations.

Question: Wells Fargo currently holds 3% of its risk weighted assets in cash and common stock, and another 2% in preferred stock. Is this ratio compliant with federal regulations?‍

Answer: Even though it's equity, preferred stock is usually classified as Additional Tier 1 (AT1) capital because it lacks full loss-absorbing capacity like common equity. In this situation, Wells Fargo would not be conforming to Basel III regulations that mandate at least 4.5% of it's assets in CET1 capital.

There’s two important features of generating data this way.

1. It’s highly accurate because we're seeding the correct answer
2. At scale, it approximates expertise. Answering the two questions above doesn’t mean your AI is an expert, but what about answering 1000 of these types of questions? What if it answered nearly perfectly for a well distributed, complex set of 10,000 of these questions? 

And that’s it! Early customer teams have reported that data generated this way is as realistic and accurate as samples they could write themselves.

Currently this requires a bit of setup as every knowledge base is different, but we threw together a basic, stripped down MVP here that you can play with – check it out here!

Note that there are some serious limitations to this web demo. For example, we haven’t exposed the underlying config file – most of our customers see massive quality jumps after just a few config iterations, and this usually fixes the first few rounds of feedback. This felt too complicated to throw up as a toy demo.

‍

[1] This is not actually how it’s phrased in Dodd Frank / Basel III but none of us want to read that today.

‍