Trained Ollama model for Beancount Query Language

80 views
Skip to first unread message

Patrick Ruckstuhl

unread,
Sep 18, 2026, 7:39:48 PMSep 18
to bean...@googlegroups.com
Hi,


I played around with training/fine-tuning a model with specific
beancount query language know-how.


I used qwen2.5-coder:7B as base and then trained it with specific
know-how based on the beanquery source code, the documentation and some
additional inputs (I used claude here to help me getting to a good test
data set).


If someone wants to play with it. Details how to pull and use are here


https://ollama.com/tarioch/qwen2.5-coder-bql


Let me know if it works for you.


Regards,

Patrick

Martin Michlmayr

unread,
Sep 19, 2026, 4:36:57 AMSep 19
to 'Patrick Ruckstuhl' via Beancount
* 'Patrick Ruckstuhl' via Beancount <bean...@googlegroups.com> [2026-09-19 01:39]:
> I played around with training/fine-tuning a model with specific beancount
> query language know-how.
>
> I used qwen2.5-coder:7B as base and then trained it with specific know-how
> based on the beanquery source code, the documentation and some additional
> inputs (I used claude here to help me getting to a good test data set).

Maybe this is off-topic here but do you have a description on how you
fine-tuned this model? What are the exact steps, what are the data
sets, etc.
--
Martin Michlmayr
https://www.cyrius.com/

Patrick Ruckstuhl

unread,
Sep 19, 2026, 5:22:42 AMSep 19
to bean...@googlegroups.com
Sure. Let me first start that I did this because I wanted to play around
with fine-tuning and found this something I can use, so this is all
still pretty new for me.


I used claude to prepare the data, I had it start with the beanquery
source code and then added old doc and also some cases and adjustments,
what was produced is now public here


https://github.com/tarioch/bql-lora


Then I used unsloth on a 12GB VRAM local graphic card to tune the model,
I had a couple attempts until I was happy enough with the state, but as
I (and maybe others) use it, I'm sure there will be some more
adjustments needed to training data and then doing another training.


Regards,

Patrick

Martin Blais

unread,
Sep 19, 2026, 4:05:29 PMSep 19
to bean...@googlegroups.com
This is cool. Added to contribs index.
Thank you for sharing,


--
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 visit https://groups.google.com/d/msgid/beancount/633e5612-2262-4266-9c30-1e5d0ea9863c%40ch.tario.org.
Reply all
Reply to author
Forward
0 new messages