Domain fine-tunes, next to the checkpoint they came from
Two QLoRA fine-tunes of Qwen2.5-1.5B for the
Adaption AutoScientist Challenge Part 2.
Every answer below was produced with the same prompt, format and greedy decoding —
the only difference is the weights. This is the complete judged evaluation, not a selection:
all 68 prompts per track, including the ones we lose.