matt-wisdom/qwen_farm_agent_gguf

FarmHand AI: Qwen 2.5 3B Instruct (Q4_K_M GGUF)

Quantized model release for FarmHand AI, an offline agricultural advisory and livestock management system developed for the Africa Deep Tech Challenge 2026 (ADTC 2026): The Laptop LLM Challenge.

---

Model Summary

• Base Model: Qwen/Qwen2.5-3B-Instruct
• Quantization Method: 4-bit medium quantization (Q4_K_M) using llama.cpp
• File Name: qwen2.5-3b-instruct.Q4_K_M.gguf
• File Size: ~1.93 GB
• Context Length: 4096 tokens
• Primary Languages: English, Nigerian Pidgin (pcm), and Hausa (ha)
• Domain Focus: Smallholder livestock management, veterinary symptom triage, least-cost feed ration formulation, and flock ledger accounting.

---

Target Hardware and Systems Profile

This model is configured specifically for resource-constrained laptops with 8 GB of total RAM and no dedicated GPU (ADTC Standard Laptop Profile).

Metric	Measured Value	Standard Limit / Baseline
Peak Memory (RSS)	~2.3 GB – 3.3 GB	7.0 GB Hard Maximum
Throughput (CPU)	16.8 Tokens/sec	15.0 Tokens/sec ($TPS_{\text{ref}}$)
Time to First Token	~180 ms	5,000 ms
Peak Core Temperature	< 65 °C	85 °C (Thermal Throttle Threshold)
Host System	Intel Core i5/i7 (2 CPU threads)	Integrated Graphics (Zero VRAM)

---

Intended Uses

1. Clinical Veterinary Triage:
   • Outbreak symptom identification (Peste des Petits Ruminants / Goat Plague, African Swine Fever, Newcastle Disease, Coccidiosis, Enterotoxemia).
   • Emergency biosecurity quarantine procedures and supportive care steps.
2. Constrained Feed Ration Formulation:
   • Extracting and validating nutritional parameters for 22 Nigerian feedstuffs (maize, palm kernel cake, soybean meal, wheat offal, bone meal, rice bran, fish meal).
3. Flock Ledger and Accounting:
   • Parsing natural language event logs (births, deaths, sales, feed purchases) into structured schema calls for SQLite persistence.
4. Multilingual Field Communication:
   • Direct conversational interaction in authentic Nigerian Pidgin and agricultural Hausa.

---

Out-of-Scope and Safety Limitations

• Not a Replacement for Licensed Veterinary Surgery: The model provides emergency first aid, symptom triage, and biosecurity checklists. It does not prescribe surgical interventions or replace a licensed field veterinarian for controlled pharmaceuticals.
• Human Clinical Medicine: The model is tuned strictly for agriculture and livestock. It must not be used for human clinical diagnosis.
• Extreme Weather Prediction: The model relies on local RAG and sensor telemetry for microclimate conditions; it does not replace national meteorological forecasting services.

---

Quickstart & Usage Examples

1. Running with llama.cpp (CLI)

./llama-cli \
  -m qwen2.5-3b-instruct.Q4_K_M.gguf \
  -p "<|im_start|>system\nYou are FarmHand AI, an on-device agricultural assistant.<|im_end|>\n<|im_start|>user\n4 goats died sudden-sudden and foam dey commot their mouth. Wetin fit cause am?<|im_end|>\n<|im_start|>assistant\n" \
  -n 256 \
  -t 2 \
  --temp 0.2

2. Running with Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="qwen2.5-3b-instruct.Q4_K_M.gguf",
    n_ctx=4096,
    n_threads=2,
    n_gpu_layers=0,  # Pure CPU execution
    verbose=False,
)

prompt = "Formulate a balanced broiler starter feed using local Nigerian ingredients with minimum 22% crude protein."

response = llm(
    f"<|im_start|>system\nYou are FarmHand AI, an on-device agricultural assistant.<|im_end|>\n<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n",
    max_tokens=256,
    temperature=0.2,
    stop=["<|im_end|>", "<|im_start|>"],
)

print(response["choices"][0]["text"].strip())

---

Repository and Codebase

• Challenge: Africa Deep Tech Challenge 2026
• Track: Agriculture
• License: Apache 2.0