15juneee/agriculture-advisor-multilingual

Agriculture Advisor Multilingual (agri_ml)

Fine-tuned for agricultural advisory in English, Swahili, Amharic, Hausa and Hindi, trained with Adaption AutoScientist for the AutoScientist Challenge (Part 2).

• Base model: meta-llama/Llama-3.2-3B-Instruct
• Training data: 15juneee/agriculture-advisor-adapted-multilingual-v1 (also on Kaggle)
• Method: AutoScientist co-optimised data adaptation and training recipe

Measured improvement

AutoScientist reported best_win_rate = 0.4851 over 5 iterations against meta-llama/Llama-3.2-3B-Instruct. That is below the 0.50 break-even point, so this model does not improve on its baseline in English - see limitations.

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 use and limitations

Intended for agricultural advisory assistance across English, Swahili, Amharic, Hausa and Hindi. 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.

Stated plainly: on AutoScientist's own evaluation this model scored a 0.4851 win rate against its base - it did not beat the baseline. About 40% of its training rows are non-English, which dilutes performance on an English-judged benchmark. It is released as a multilingual-capability artifact, not as an English-performance improvement; prefer the English-only sibling model where English quality is what matters.

Reproducing

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