Professor/naija-agri-llm-yoruba-hausa

Naija Agri LLM — Yoruba & Hausa Agricultural Assistant

A LoRA adapter for Llama 4 Scout 17B 16E Instruct (109B total parameters, MoE) fine-tuned on Nigerian agricultural Q&A data in Yoruba and Hausa using AutoScientist by Adaption Labs.

Built for the Adaption Labs AutoScientist Challenge (Agriculture category, 2026) — powering a crosslingual RAG-based agricultural assistant for Nigerian smallholder farmers.

What It Does

Nigerian farmers and agricultural extension workers can ask questions in Yoruba, Hausa, or English and receive accurate, actionable answers grounded in Nigerian farming context — referencing institutions like IITA, NAERLS, and state ADP offices, and Nigerian crop varieties (cassava, yam, maize, cowpea, cocoa, plantain, tomatoes).

This model is the generative backbone of a RAG pipeline that supports crosslingual retrieval: a farmer's question in Yoruba retrieves relevant document chunks, which are passed as context to this model for answer generation in the farmer's language.

Training

Parameter	Value
Base model	meta-llama/Llama-4-Scout-17B-16E-Instruct (109B total params)
Training infrastructure	Together AI (via AutoScientist)
Training method	SFT + LoRA
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.0
Target modules	All linear layers
Epochs	2
Batch size	Max
Learning rate	2e-4 (cosine decay, min ratio 0.1)
Warmup ratio	0.05
Max grad norm	1
Evaluations	5 checkpoints

Training Metrics

Metric	Value
Starting train loss	~1.93
Final train loss	~0.94
AutoScientist Win Rate (adapted)	64%
AutoScientist Win Rate (base)	36%

The adapted model wins 64 out of 100 head-to-head comparisons against the base Llama 4 Scout on agricultural Q&A in Yoruba and Hausa, per AutoScientist's internal preference evaluation. Train and validation loss converge smoothly with no signs of overfitting across 94 steps.

Dataset

Trained on Professor/agronomy-qa-pairs (~35,900 rows), localized from KisanVaani/agriculture-qa-english-only using Adaption Labs Data Adaption with a Nigerian context blueprint (IITA, NAERLS, ADP).

Language distribution:

• Yoruba (yo): ~30,900 rows
• Hausa (ha): ~4,600 rows
• Nigerian English (en): ~360 rows

Intended Use

• Yoruba and Hausa speaking smallholder farmers in Nigeria
• Agricultural extension workers providing guidance in local languages
• RAG pipelines for crosslingual agricultural document retrieval
• Research on multilingual LLM adaptation for low-resource African languages

Languages

This model responds in Yoruba (with full tone marks: ọ, ẹ, ṣ, etc.), Hausa (standard orthography), and English. The English capability comes from both the base model and Nigeria-specific English training rows, so English responses reflect Nigerian agricultural context rather than generic advice. It deflects out-of-scope questions in the user's language.

Deployment

This adapter was trained on Together AI infrastructure. The trained model name on Together AI is adaption_agronomy_qa_pairs. It is compatible with Together AI's fine-tuned model serving — the base Llama 4 Scout 17B is hosted on Together AI and can be served with this LoRA.

Limitations

• Fine-tuned on agriculture domain only; will deflect general queries
• Hausa coverage is smaller (~13% of training data) — may be less robust than Yoruba
• Quality of Yoruba tone marks depends on the localization pipeline used

License

Built on Llama 4 — subject to Meta's Llama 4 Community License Agreement.