Interesting idea! Does it have a special name? So, effectively we have options:
- Train ML model on our historical data and historical classifications - Ask existing LLM to classify transactions (some privacy issues here) - Ask LLM to write a deterministic code to classify (your example) On Monday, March 11, 2024 at 2:28:39 PM UTC+1 [email protected] wrote: > Yup, you nailed it! That's the idea :) it just saves time from worrying > about building a machine learning model, because we all have access to a > big one in the form of a large language model (Google/OpenAI are paying the > electricity cost). > > So this can help save time by removing the overhead and creating a > "deterministic" classifier in the form a python program that is likely more > concise, maintainable, and interpretable than machine learning models > trained for this task. > > > > On Mon, Mar 11, 2024 at 7:03 AM Chary Chary <[email protected]> wrote: > >> Hi, >> >> I briefly scanned through the article. >> >> So in the essence you give LLM a bunch some CSV example and then ask LLM >> to write a python code, which would categorize similar, based on the >> keywords. >> >> So, this is a kind of alternative to teaching ML model based on the >> previous transactions to be able to categorize new ones. >> >> Did it get the main idea correctly? >> >> On Monday, March 11, 2024 at 12:36:00 AM UTC+1 [email protected] wrote: >> >>> Made a quick prompt over the weekend: >>> https://gist.github.com/jaanli/1f735ce0ddec4aa4d1fccb4535f3843f >>> >>> Results are that my partner (someone non-technical, design background, >>> but familiar with prompt engineering) can use the prompts—the last thing I >>> would want is an inscrutable system that I manually built to import >>> transactions from our dozen institutions across multiple countries & >>> currencies, that they can't re-use or extend. >>> >>> Visual Studio Code and the Beancount extension are already a stretch for >>> them so having something that works with a single prompt at a time and copy >>> and pasting was my goal. >>> >>> Hope this helps someone else! Surprised that these tools are not easier >>> to use (and thank you for beancount, this wouldn't be possible otherwise :) >>> >>> Would be fun to extend this with DSPy ( >>> https://github.com/stanfordnlp/dspy/blob/main/intro.ipynb) which could >>> likely help squeeze several different converters into a few signatures >>> (compressed prompts), and things like chain-of-thought prompting (iterative >>> runs of large language models) would further reduce the >>> extract-transform-load overhead that has kept me from trying beancount all >>> these years. >>> >>> Very best, >>> Jaan >>> >> -- >> You received this message because you are subscribed to a topic in the >> Google Groups "Beancount" group. >> To unsubscribe from this topic, visit >> https://groups.google.com/d/topic/beancount/aoZ7-H1tCX4/unsubscribe. >> To unsubscribe from this group and all its topics, send an email to >> [email protected]. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/beancount/eefe5ba4-c0cd-46cd-9fb8-8922b399c221n%40googlegroups.com >> >> <https://groups.google.com/d/msgid/beancount/eefe5ba4-c0cd-46cd-9fb8-8922b399c221n%40googlegroups.com?utm_medium=email&utm_source=footer> >> . >> > -- You received this message because you are subscribed to the Google Groups "Beancount" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/beancount/45eb3d9e-fd7c-4ad5-a36b-f9d3f3ac411fn%40googlegroups.com.
