qwen25_15b_goldadtc_submission

ARIS: Offline Agricultural AI for Nigerian Smallholder Farmers 🌾

ARIS (Agricultural Recommendation and Information System) is a highly specialized, 1.5B-parameter Small Language Model (SLM) fine-tuned for offline deployment on resource-constrained hardware. It is submitted as part of the Africa Deep Tech Community (ADTC) 2026 Hackathon (Gate 2 Semifinals).

This repository contains the Q4_K_M quantized GGUF weights, heavily optimized to run locally on an 8GB RAM CPU-only laptop without internet access, live weather feeds, or camera inputs.

📊 Model Details

• Base Model: Qwen/Qwen2.5-1.5B-Instruct
• Architecture: QLoRA Fine-tune (Rank = 64, Alpha = 128)
• Quantization: GGUF Q4_K_M (via llama.cpp)
• File Size: ~986 MB (Highly efficient footprint leaving ~7GB RAM free for the host OS)
• Developer: Victor Chukwuebuka Nwaruwe (Vicgrace Labs)

📚 Training Data & Provenance

ARIS is not a general-purpose chatbot. It was strictly fine-tuned on a custom, adversarially-curated dataset comprising over 1,000 verified agricultural records.

• Core Sources: Official production manuals and bulletins from NAERLS, IITA, NRCRI, NIHORT, NSPRI, NAPRI, NVRI, and NiMet.
• Language Support: Fluent in English and Nigerian Pidgin, maintaining strict agronomic accuracy across language switching.
• Adversarial Curation: The dataset features "Contrastive Boundary" examples to prevent Concept Bleed (e.g., separating Maize and Yam spacing logic) and anti-sycophancy training (teaching the model to confidently reject dangerous but confident prompts from users).

🛡️ Safety & Alignment Constraints

To prevent dangerous agricultural hallucination, ARIS was trained with a "Zero-Trust" refusal policy:

1. Chemical/Fertilizer Dosages: Explicitly refuses to provide flat NPK or pesticide dosages without a confirmed soil test, product label, and sprayer calibration.
2. Veterinary & Medical: Refuses medical diagnoses or veterinary prescriptions, directing users to qualified professionals or emergency services.
3. Capability Honesty: Openly acknowledges its inability to see uploaded photos, fetch live market prices, or predict live weather (limiting climate advice to NiMet seasonal predictions).

💻 Example Usage (llama.cpp)

You can run this model locally on any standard CPU using llama.cpp.

Standard CLI Execution:

./llama-cli -m qwen2.5-1.5b-instruct.Q4_K_M.gguf \
  -p "<|im_start|>system\nYou are ARIS, an offline agricultural advisor for Nigerian smallholder farmers.\n- No internet access\n- No camera\nAnswer strictly from verified Nigerian agricultural sources. When the user writes in Nigerian Pidgin, respond in Nigerian Pidgin.<|im_end|>\n<|im_start|>user\nWetin be the correct way to plant yam for rainy season?<|im_end|>\n<|im_start|>assistant\n" \
  -n 256 \
  --temp 0.1