Ask AgriLM
Our own 1.5-billion-parameter agricultural language model, trained from scratch. Each answer is checked against trusted farming guides before you see it.
AgriLM-1.3B
A small language model for African agriculture that we trained from scratch — no borrowed weights, no third-party model API. Open, Llama-compatible architecture: it runs on our servers, on a laptop, or offline.
Pretraining, live
Loss on text the model has never seen. Lower is better.
Versions
Every version of AgriLM we have trained, oldest first. Pick one in Ask to talk to it, or compare answers with “Try another version”.
What training changed: same question, two stages
The base model has only read agricultural text (pretraining). AgriLM is the same network after we fine-tuned it on expert Q&A. No reference library is used here — this is the model alone.
Measured quality
Held-out tests the model never saw in training.
Architecture
What it read
Run it anywhere
Exported to the open GGUF format. Works with llama.cpp, Ollama and LM Studio — no internet, no GPU required.
Speed test on this machine
Make AgriLM yours
Upload your documents and expert answers. AgriLM learns your domain, your language and your way of advising — and you get a new model version you own.
Documents or Q&A pairs
We validate, de-duplicate and count tokens
Continued pretraining, fine-tuning, or both
Loss on held-out examples, logged per epoch
New version, also as a portable GGUF file
📚 Continued pretraining
For large text collections — reports, extension archives, local-language text. Teaches the model new vocabulary and background knowledge.
Needs ~10M+ tokens to matter. Mixed with our original corpus so the model doesn't forget.
🎯 Fine-tuning
For question–answer pairs — expert answers, call-centre logs. Teaches how to answer: tone, format, when to refer to an officer.
500+ pairs recommended. Mixed with our 38k general examples.
⚡ Reference library
For a few documents you need cited today. No training — searchable in seconds.
1 · Upload training data
2 · Your datasets
3 · Start a training job
4 · Jobs
Connect your knowledge
Documents you add are searchable straight away, and AgriLM checks its answers against them. For deeper changes to the model itself, use Train.
Upload documents
Add a web page
Fetches that one page and adds its text. Respect the site's terms of use.
Library
Connectors
277 East African farming guides (CC BY-NC-SA).
1,000 expert-curated answers.
Via the C4IR open API / MCP server once access is granted.
API
Connect AgriLM to any channel — web, WhatsApp, SMS gateways, extension apps. JSON over HTTPS, with cited sources and a request ID for every answer.
API keys
POST /v1/ask
Question in, answer with sources out.
{ "request_id": "…", "answer": "…", "language": "en",
"sources": [{"title": "…", "source": "https://…"}],
"checks": {"mode": "generated", "corrected_sentences": 0,
"unverified_sentences": 0},
"model": "agrilm-1.3b-v2/sft_v5/epoch_2" }
POST /v1/chat/completions
OpenAI-compatible — works with existing OpenAI client libraries.
Try it
Response appears here.
What farmers are asking
Every answer is logged with its sources and checks — the audit trail behind each response, and a live view of farmers' needs. Clients are pseudonymised.