ChatGPT & Beancount

[Martin Blais]

It's a simplistic example, but still... if you haven't had a try at GPT to generate code, look what it can do with importers, not a bad way to get started:

[Chao Chen]

It's really cool! ChatGPT could also be a great account classifier for beancount importers. The only problem is privacy...

On Fri, 31 Mar 2023 at 09:16, Martin Blais <bl...@furius.ca> wrote:

It's a simplistic example, but still... if you haven't had a try at GPT to generate code, look what it can do with importers, not a bad way to get started:

--
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 beancount+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/beancount/CAK21%2BhO5Xv-sFWBNTm9MdB6AGfgdQLhy3A9LM9BnGBWyJR-qmQ%40mail.gmail.com.

[Red S]

Fantastic! Hooray for reducing drudgery!

It's apparently quite effective at using code from existing github projects. I love the code explanations in natural language.

[Martin Blais]

I haven't tried yet but my hope is that gpt would be great at writing test code. Can you imagine?

On Fri, Mar 31, 2023, 01:03 Red S <redst...@gmail.com> wrote:

Fantastic! Hooray for reducing drudgery!

It's apparently quite effective at using code from existing github projects. I love the code explanations in natural language.

--

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 beancount+...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/beancount/d7116b04-c77d-450d-b015-2f2148e9664en%40googlegroups.com.

[Red S]

Great idea. Talk about drudgery!

[Reed Law]

I used a command-line tool, aichat, to pipe transactions to GPT-4 and categorize expenses. I gave it the following input (generated according to Booking Amazon Purchases):

2022-08-19 * "Amazon.com" "Beginning Drawing Atelier: An Instructional Sketchbook"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -24.71 USD

2022-09-14 * "Amazon.com" "Amazon Basics Clear Sheet Protectors for 3 Ring Binder, 8.5 x 11 Inch, 200-Pack"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -18.16 USD

2022-10-05 * "Amazon.com" "Bosch Rear Wiper Blade H307 /3397011429 Original Equipment Replacement- 12\" (Pack of 1)"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -11.6 USD

2023-01-26 * "Amazon.com" "Duxtop 1600W Single Burner Electric Cooktop, Electric Hot Plate for Cooking, Electric Stove with Sensor Touch Control, Portable Infrared Burner with T"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -69.43 USD

2023-01-26 * "Amazon.com" "PILOT Parallel Caligraphy Pens, Assorted Point Sizes and Colors, 6 Count (12722)"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -32.44 USD

2023-02-28 * "Amazon.com" "Laptop Stand, Laptop Holder Riser Computer Stand, Adjustable Aluminum Foldable Portable Notebook Stand, Compatible with MacBook Air Pro, HP, Lenovo, D"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -14.29 USD

Using the following prompt and command:

cat amz.beancount | aichat add categorized expense postings to balance each of the following transactions, preserving the included zero sum posting. do not use spaces in account names > out.beancount

I got these results:

2022-08-19 * "Amazon.com" "Beginning Drawing Atelier: An Instructional Sketchbook"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -24.71 USD
  Expenses:Art:Instructional-Sketchbook

2022-09-14 * "Amazon.com" "Amazon Basics Clear Sheet Protectors for 3 Ring Binder, 8.5 x 11 Inch, 200-Pack"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -18.16 USD
  Expenses:Office-Supplies:Sheet-Protectors

2022-10-05 * "Amazon.com" "Bosch Rear Wiper Blade H307 /3397011429 Original Equipment Replacement- 12\" (Pack of 1)"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -11.6 USD
  Expenses:Car:Maintenance

2023-01-26 * "Amazon.com" "Duxtop 1600W Single Burner Electric Cooktop, Electric Hot Plate for Cooking, Electric Stove with Sensor Touch Control, Portable Infrared Burner with T"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -69.43 USD
  Expenses:Kitchen:Appliances

2023-01-26 * "Amazon.com" "PILOT Parallel Caligraphy Pens, Assorted Point Sizes and Colors, 6 Count (12722)"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -32.44 USD
  Expenses:Art:Calligraphy-Pens

2023-02-28 * "Amazon.com" "Laptop Stand, Laptop Holder Riser Computer Stand, Adjustable Aluminum Foldable Portable Notebook Stand, Compatible with MacBook Air Pro, HP, Lenovo, D"
  Assets:Zero-Sum-Accounts:Amazon-Purchases  -14.29 USD
  Expenses:Office-Supplies:Laptop-Stand

There was some trial and error that resulted in the final prompt. At first, it removed the zero-sum postings and only output expenses. The expense accounts had spaces in their names (e.g. "Expenses:Office Supplies:Laptop Stand"). There is still room for improvement (I don't really care to have an account for laptop stands). Overall, it looks very helpful. Perhaps a good workflow is to use GPT to categorize a small batch of expenses, manually correct them, and then use smart_importer to handle the rest.

If anyone else has tips please share!

[Martin Blais]

Mind blown *poofff*

--

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 beancount+...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/beancount/38f90ec4-fea3-4794-98de-4d90f28c960dn%40googlegroups.com.

[Red S]

Very neat! I've got to admit, I'm bowled over by the natural language interface working so well.

Amazon purchase classification is almost the only area of classification where smart_importer doesn't do greatly well for me, since there are few repeat purchases or keywords.

Good to know about aichat too.

That said, here are the results I tried to replicate with the same data. Note the claim at the bottom: "You can use any account name you like as long as it doesn’t contain spaces," and its incongruency with literally every classification containing a space. One of the limitations of the current generation is a lack of a confidence-in-results indicator, which shows up here.

2022-08-19 * "Amazon.com" "Beginning Drawing Atelier: An Instructional Sketchbook" Assets:Zero-Sum-Accounts:Amazon-Purchases -24.71 USD Expenses:Art Supplies 2022-09-14 * "Amazon.com" "Amazon Basics Clear Sheet Protectors for 3 Ring Binder, 8.5 x 11 Inch, 200-Pack" Assets:Zero-Sum-Accounts:Amazon-Purchases -18.16 USD Expenses:Office Supplies 2022-10-05 * "Amazon.com" "Bosch Rear Wiper Blade H307 /3397011429 Original Equipment Replacement- 12\" (Pack of 1)" Assets:Zero-Sum-Accounts:Amazon-Purchases -11.6 USD Expenses:Auto Parts 2023-01-26 * "Amazon.com" "Duxtop 1600W Single Burner Electric Cooktop, Electric Hot Plate for Cooking, Electric Stove with Sensor Touch Control, Portable Infrared Burner with T" Assets:Zero-Sum-Accounts:Amazon-Purchases -69.43 USD Expenses:Kitchen Appliances 2023-01-26 * "Amazon.com" "PILOT Parallel Caligraphy Pens, Assorted Point Sizes and Colors, 6 Count (12722)" Assets:Zero-Sum-Accounts:Amazon-Purchases -32.44 USD Expenses:Art Supplies 2023-02-28 * "Amazon.com" "Laptop Stand, Laptop Holder Riser Computer Stand, Adjustable Aluminum Foldable Portable Notebook Stand, Compatible with MacBook Air Pro, HP, Lenovo, D" Assets:Zero-Sum-Accounts:Amazon-Purchases -14.29 USD Expenses:Laptop Accessories

In this example, I’ve added a category to each transaction by adding a new line after the transaction line and indenting it by two spaces. The category is specified as an account name and is preceded by a colon. You can use any account name you like as long as it doesn’t contain spaces.

[Martin Blais]

FYI for those who might be tempted to feed their entire ledger into the thing, keep in mind the privacy implications of doing this.

The queries are very likely logged and even if used indirectly as training data for further updates, might leak your personal info to someone else generating transaction data. 

Just saying... it might not be obvious to some.

--

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 beancount+...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/beancount/b08ca160-539a-4d7a-ba9c-8540f214d2bcn%40googlegroups.com.

[Reed Law]

Good reminder about privacy. There is an open source chatbot, Vicuna, which plans to release training weights next week meaning you can use it privately and offline.

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/_7udAMlfLAg/unsubscribe.
To unsubscribe from this group and all its topics, send an email to beancount+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/beancount/CAK21%2BhOGB19cwNOcyBmvEbZAcwNCEnUuN-k-io3ifbkZe3vheA%40mail.gmail.com.

[Reed Law]

I got decent results with this script. It uses the LangChain Python library with LLaMA.cpp and the Alpaca LLM model. The results aren't as good as with the OpenAI API but it keeps everything local and private.

[Reed Law]

This version uses an OpenAI API key and transforms the CSV file adding a Category column.

[Bhishmaraj S]

Has anyone tried finetuning (LoRA) on smaller LLMs for this task?