v1.1.0

Changes

🚀 Features

• Return a loose score next to the strict one from the evaluation check (#19) @tarioch
• Consistent answer shapes, query-only answers and a larger evaluation set (#18) @tarioch
• Use the llm-finetune skills for training and serving (#14) @tarioch, Patrick Ruckstuhl

🐛 Bug Fixes

• Score an answer with a string where an amount is expected as wrong (#16) @tarioch
• Keep every example of a prompt on one side of the train/val split (#15) @tarioch, @claude

🧰 Maintenance

• Update README training numbers to the measured ones (#17) @tarioch

v1.0.0

First published release of the training data and the models built from it.

Published on Ollama as tarioch/qwen2.5-coder-bql:

Tag	Base model	Size (Q4_K_M)
:7b	Qwen2.5-Coder-7B-Instruct, QLoRA rank 32	4.4 GiB
:3b (= latest)	Qwen2.5-Coder-3B-Instruct	1.8 GiB

Both were trained on data/ as of this release. Compared with the previous models, they handle:

• the last entry per account among accounts active in a window ("... that had an entry this year, but older than 2 months"), using HAVING max(date) < today() - interval('2 months')
• rolling windows such as "in the last 3 months", using today() - interval('3 months')

Run ollama pull tarioch/qwen2.5-coder-bql:7b (or :3b) to update.