agripadi-tiny-aya-earth-q4_k_m

AgriPadi Tiny-Aya Earth Q4_K_M

A QLoRA-fine-tuned quantized version of CohereLabs/tiny-aya-earth (3B, Cohere2ForCausalLM) targeting smallholder agriculture advisory in West Africa.

What This Model Is

• Base: CohereLabs/tiny-aya-earth — a 3B multilingual model covering 100+ languages including Yoruba, Igbo, Hausa, and Nigerian Pidgin.
• Fine-tuned on AgriPadi SFT data: multiple-choice agriculture questions, short-answer Q&A, and conversational advisory pairs on crop management, disease diagnosis, soil health, and pest control.
• Quantized to Q4_K_M (~2.0 GB) via standalone llama.cpp (patched converter for Cohere2 BPE fingerprint 89377a9f...).

Training Details

Training ran on an NVIDIA A10 (24 GB) via Modal.

What It Can Do

• Answer multiple-choice and short-answer agriculture questions (English, Yoruba, Pidgin)
• Diagnose crop diseases from descriptions
• Recommend fertilizer application rates and timing
• Explain soil health practices for Nigerian farming contexts
• Conversational advisory in English, Yoruba, Hausa, Igbo

What It Cannot Do

• Process images (text-only; no vision encoder)
• Replace professional agronomic advice for high-stakes decisions
• Reason reliably on topics outside its training distribution
• Handle very long contexts beyond 2048 tokens

Usage

llama.cpp

# download
huggingface-cli download Alabi-Ayobami/agripadi-tiny-aya-earth-q4_k_m \
    model-q4_k_m.gguf --local-dir .

# run
./llama-cli -m model-q4_k_m.gguf \
    -p "What are the signs of nitrogen deficiency in maize?" \
    -n 512 -ngl -1

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(model_path="model-q4_k_m.gguf", n_ctx=2048, n_gpu_layers=-1)

prompt = "User: How do I control cassava mealybug?\nAnswer:"
out = llm.create_completion(prompt, max_tokens=512, temperature=0.0)
print(out["choices"][0]["text"])

HuggingFace Hub

from huggingface_hub import hf_hub_download
from llama_cpp import Llama

path = hf_hub_download(
    "Alabi-Ayobami/agripadi-tiny-aya-earth-q4_k_m",
    filename="model-q4_k_m.gguf",
)
llm = Llama(model_path=path, n_ctx=2048)

Quantization

Q4_K_M uses 4-bit quantization with K-quant mixed scales — good balance of speed and accuracy for a 3B model. The quantizer is llama-quantize from llama.cpp (built from source with tiny_aya fingerprint patch).

Training Data

• MC (multiple-choice): ~3,700 agriculture questions with 5 options each
• Short answer: ~600 open-ended Q&A pairs on crop management, livestock, soil
• Advisory: ~700 conversational pairs (farmer question → expert answer)
• Languages: primarily English, with Yoruba, Hausa, Igbo, and Nigerian Pidgin samples

Limitations

• Small base model (3B): less capable than larger models on complex reasoning
• Domain-specific fine-tune: may produce confident but incorrect answers outside agriculture
• Quantized: Q4_K_M introduces minor precision loss vs. F16; critical applications should verify against the full-precision adapter
• Validation gap: evaluated on general knowledge (ARC-Easy), not a domain-specific agriculture benchmark

Citation

@misc{agripadi2026,
  title={AgriPadi: Smallholder Agriculture Advisory with Fine-Tuned Language Models},
  author={Alabi Ayobami},
  year={2026},
  howpublished={\url{https://huggingface.co/Alabi-Ayobami/agripadi-tiny-aya-earth-q4_k_m}}
}

License

Apache 2.0. Inherits the base model's license (CohereLabs/tiny-aya-earth).