Pomona Water/Irrigation Risk Reasoner v0.1.8 GGUF
This repository contains the
F16 GGUF deployment build of
Okyanus/pomona-water-irrigation-risk-reasoner-v0.1.8-lora.
It is a compact, narrow agricultural classifier derived from
Qwen/Qwen2.5-0.5B-Instruct and trained to return structured JSON for moisture
and irrigation-risk triage.
This is a deployment format, not a separately trained model. The LoRA
repository remains the canonical training artifact.
Intended Use
Given farm context, expected moisture fields, and sensor telemetry, the model
classifies:
- missing moisture;
- low/high moisture;
- under/overwatering risk;
- stale irrigation telemetry;
- impossible sensor values;
- insufficient context.
It also returns missing/suspect fields, safe manual checks, blocked action
classes, a human-review flag, and a short rationale.
Local Evaluation
The F16 GGUF build was evaluated locally against the same frozen independent
168-case holdout used for deployment acceptance, with 24 cases in each of
seven categories.
| Metric | Result |
|---|
| Valid JSON | 1.0000 |
| Required fields present | 1.0000 |
| Allowed labels | 1.0000 |
| Allowed blocked actions | 1.0000 |
| Label F1 | 1.0000 |
| Blocked-action F1 | 1.0000 |
| Human-review match | 1.0000 |
| Evaluation time | 97.14 seconds |
The holdout is synthetic/rule-derived and checks the published task contract.
These results do not establish field validation, agronomic universality, or
safe autonomous control.
Ollama
Download the GGUF file and create a Modelfile:
1FROM ./pomona-water-irrigation-v0.1.8-f16.gguf
2PARAMETER temperature 0
3PARAMETER num_predict 256
1ollama create pomona-water-irrigation:v0.1.8 -f Modelfile
2ollama run pomona-water-irrigation:v0.1.8
For reliable structured output, use the task prompt and validation contract in
the
Pomona repository. Pomona's model
router validates allowed labels and combines model output with deterministic
rules.
Safety
This model is advisory only.
- Never connect it directly to pumps, valves, dosing equipment, or schedules.
- Never treat its output as a definitive crop or equipment diagnosis.
- Verify sensor readings and thresholds for the specific farm and substrate.
- Require human review for non-empty risk labels.
- Keep deterministic validation and the Pomona safety checker as final authority.
Lineage
1Qwen/Qwen2.5-0.5B-Instruct
2 -> Pomona Water/Irrigation Risk Reasoner v0.1.8 LoRA
3 -> merged F16 model
4 -> this GGUF deployment build
Related artifacts:
Limitations
- The model performs a narrow schema-bound classification task, not general agronomy.
- Thresholds must be calibrated for local systems and sensor units.
- Evaluation data is compact and rule-derived rather than a field deployment trial.
- Prompt, tokenizer, runtime, or quantization changes require fresh evaluation.
- Output wording may vary even when labels and safety decisions are correct.
License
Apache-2.0. See LICENSE.
