Fine-tuned for agricultural advisory, trained with Adaption AutoScientist for the AutoScientist Challenge (Part 2).

AutoScientist reported best_win_rate = 0.5711 against meta-llama/Llama-3.2-3B-Instruct - the fine-tuned model is preferred over its own baseline in 57.1% of comparisons.

Trained on the Adaption-adapted English agriculture dataset with domain augmentation (3,000 domain rows + 500 general). Augmentation was the decisive factor: the same adapted data trained without it scored exactly 0.5000 (break-even), and Adaption's own dataset evaluation scored the adaptation itself at 7.0 -> 9.0 (+28.6%).

Evaluation methodology, including the position-swap and dual-judge controls, is in EVAL.md in the project repository. The held-out split used is published alongside the training data so the number can be reproduced.

Intended for agricultural advisory assistance. Not a substitute for local agricultural extension services. Any pesticide, herbicide or veterinary guidance must be checked against the current product label and local regulations, which govern legal rates and pre-harvest intervals.

Training data covers smallholder systems in Kenya, Ethiopia, Nigeria and India plus temperate broadacre agronomy. It will be weaker outside those contexts.

The dataset build, training pipeline and evaluation harness are all scripted; see the project repository.