Okyanus/pomona-water-irrigation-risk-reasoner-v0.1.8-MLX
Pomona Water/Irrigation Risk Reasoner v0.1.8 MLX
This repository contains the accepted 8-bit MLX deployment build of Okyanus/pomona-water-irrigation-risk-reasoner-v0.1.8-lora for Apple Silicon.
This is a deployment conversion, not a separately trained model. The LoRA repository remains the canonical training artifact.
Intended Use
The model performs narrow, structured moisture and irrigation-risk triage. It returns JSON containing risk labels, missing and suspect sensor fields, safe manual checks, blocked action classes, a human-review flag, and rationale.
Local Evaluation
The 8-bit MLX build passed the frozen independent 168-case deployment holdout.
The separately tested 4-bit MLX conversion was rejected because label F1 fell to 0.7262 and human-review match to 0.5536. That rejected build is not included in this repository.
The holdout is synthetic/rule-derived and verifies the task contract. It is not field validation or evidence for autonomous irrigation control.
Run on Apple Silicon
python3 -m venv .venv
source .venv/bin/activate
pip install mlx-lm
mlx_lm.server \
--model Okyanus/pomona-water-irrigation-risk-reasoner-v0.1.8-MLX \
--port 8083The local MLX server exposes an OpenAI-compatible chat endpoint. It is a development server and is not recommended as a hardened public service.
For the exact structured prompt, output validator, and guarded API integration, see the Pomona repository.
Safety
This model is advisory only.
- Never connect it directly to pumps, valves, dosing equipment, or schedules.
- Validate every response against the allowed schema and labels.
- Require human review for non-empty risk labels.
- Keep deterministic checks and the Pomona safety checker as final authority.
- Verify thresholds, units, and sensor quality for the local farm.
Lineage
Qwen/Qwen2.5-0.5B-Instruct
-> Pomona Water/Irrigation Risk Reasoner v0.1.8 LoRA
-> merged model
-> this accepted 8-bit MLX deployment buildRelated artifacts:
Limitations
- Narrow schema-bound classifier, not a general agricultural assistant.
- Evaluation data is compact and rule-derived.
- No field-trial, causal, yield, or equipment-safety validation is claimed.
- Runtime and prompt changes can alter behavior and require reevaluation.
License
Apache-2.0. See LICENSE.
