AgriContext-IE

AgriContext-IE is an Ireland-first agricultural decision-support adapter for field trafficability, soil damage, grazing, and poaching-risk questions. It was trained with Adaption AutoScientist for the Agriculture category of the AutoScientist Challenge.

The model is designed to use the local conditions supplied in a prompt—such as drainage, recent weather, current field observations, and machinery or livestock pressure. When essential information is absent, it should return INSUFFICIENT_CONTEXT instead of guessing.

This release contains a LoRA adapter, tokenizer files, chat template, and training metadata. It does not contain the full 70B base-model weights.

Performance

Adaption AutoScientist completed three research iterations and reported a 66.19% best head-to-head win rate for the adapted model against the original base model.

Measure	Result
AutoScientist status	Succeeded
Best adapted-vs-base win rate	66.19%
Completed AutoScientist iterations	3 / 3
Final recorded evaluation loss	1.168

The win rate means the adapted model was preferred in approximately two-thirds of AutoScientist's pairwise comparisons. It is not a claim of 66.19% absolute accuracy or 66.19% relative improvement. The project's separate 20-case benchmark has not yet been scored, so no independent benchmark result is claimed here.

Intended behaviour

The model returns one of four labels:

LOW_RISK | MODERATE_RISK | HIGH_RISK | INSUFFICIENT_CONTEXT

Responses follow this structure:

RISK:
ASSESSMENT:
RELEVANT FACTORS:
MISSING INFORMATION:
NEXT STEP:
CONFIDENCE:

Primary use cases are:

• Machinery access and soil structural-damage risk
• Livestock grazing and poaching risk
• Weather-sensitive field-operation timing
• Detecting missing or overly broad agricultural context

Loading the adapter

The base model is large and may require gated access and substantial accelerator memory. Follow the base model's licence and access requirements.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_id = "togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference"
adapter_id = "agricontext-ie"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)

messages = [
    {
        "role": "system",
        "content": (
            "You are AgriContext-IE, a context-aware agricultural "
            "decision-support assistant for Irish farming conditions. "
            "Use only the conditions supplied. If essential information "
            "is missing, return INSUFFICIENT_CONTEXT."
        ),
    },
    {
        "role": "user",
        "content": (
            "Can I take a loaded tractor and trailer onto my grass field "
            "today? The nearest rain gauge recorded 18 mm over 48 hours."
        ),
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=500, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Set adapter_id to the local adapter directory or its published Hugging Face repository ID.

For consistent results, use the full system blueprint distributed with the project rather than the shortened example above.

Training details

Setting	Value
Base model recorded by AutoScientist	meta-llama/Llama-3.3-70B-Instruct-Reference
Compatible base path in adapter config	togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference
Method	Supervised fine-tuning
Adapter	LoRA / PEFT
Rank / alpha	8 / 8
Target modules	q_proj, v_proj
Epochs	1
Optimizer schedule	Cosine
Learning rate	1e-4
Warmup ratio	0.1
Maximum gradient norm	2
Precision recorded in config	bfloat16
Trainer steps	102
Evaluation events	5

The training set contains 40 reviewed English examples, evenly balanced across the four risk labels. It is deliberately narrow and focuses on Irish dairy and grassland field conditions. Agricultural claims were traced to official Met Éireann and Teagasc material during dataset construction.

AutoScientist run metadata:

• Training experiment: 963e2d81-7268-45e4-b019-85b8a2abf57a
• Fine-tuning job: 4080ab61-6ecd-4539-84ed-e328eebd62cf
• Training type: LoRA
• Training method: SFT

Limitations and safety

• The training set is small and the domain is intentionally narrow.
• The initial release is English-only and Ireland-first.
• Conditions can vary between nearby fields; regional data is not proof of a particular field's condition.
• LOW_RISK does not mean zero risk or guarantee suitability.
• The model must not invent weather readings, soil conditions, regulations, products, or dosages.
• Pesticide, veterinary medicine, and chemical-dosage instructions are outside scope.
• Output is not proof of regulatory compliance and does not replace direct inspection or advice from a farmer, agricultural adviser, agronomist, or another qualified professional.

Release contents

• adapter_model.safetensors — LoRA adapter weights
• adapter_config.json — PEFT adapter configuration
• config.json — base architecture configuration
• tokenizer.json and tokenizer metadata
• chat_template.jinja — Llama chat template
• trainer_state.json — trainer logs and evaluation losses
• autoscientist_config.json — AutoScientist run and recipe metadata

Framework versions

The exported checkpoint records PEFT 0.15.1 and Transformers 5.13.0. Newer compatible releases may also work, but were not validated as part of this release.