Sreekanth Harigovindan PRO

Good call, will try it.

Fair point on the dataset. The 202 pairs are just requirement→test — the page itself isn't baked into the model. It gets added when you actually run it: the tool grabs the live page via --url and hands that to the model right before it writes the test. So the model learns the "turn a requirement into a test" part, and the page details come in at runtime — same as the cloud version.

Source → https://github.com/aiqualitylab/ai-natural-language-tests

Curious how more reasoning ability handles the tricky pairs. If you train something, please share it and I will link it from the collection. Thanks!

Thanks, I went with Qwen2.5-Coder-1.5B to keep it lightweight (~1.6 GB) and fully offline, but trying a newer base is on my list. Dataset's open.