Pomona Model Catalog¶
This is Pomona's GitHub-facing model card index. It records model family status, lineage, runtime formats, and safety boundaries. Weights are not stored in this repository.
All published models and the dataset are also collected in one place on Hugging Face: Pomona — Local AI for Safer Greenhouse Decision Support.
Try the guarded tomato reasoner without installing anything: Pomona Greenhouse Demo.
Current Families¶
| Family | Version | Status | Canonical artifact | Runtime role |
|---|---|---|---|---|
| Agronomist assistant | Gemma4 | Published | Okyanus/ai-pomona-agronomist-gemma4 | Broad advisory explanation |
| Tomato risk | v0.1.7 | Published | Okyanus/pomona-tomato-risk-reasoner-v0.1.7-lora | Tomato greenhouse risk labels |
| Water / irrigation risk | v0.1.8 | Published release candidate | Okyanus/pomona-water-irrigation-risk-reasoner-v0.1.8-lora | Moisture and irrigation triage |
| Actuator command gate | v0.1 | Published research preview | Okyanus/pomona-actuator-command-gate-reasoner-v0.1-lora | Advisory command classification; deterministic gate is final |
| Sensor quality | v0.1.1-boundary | Unpublished local candidate | Local adapter only | Telemetry quality classification |
| Safety triage | v0.1 | Unpublished local candidate | Local adapter only | Action safety classification |
| Nutrient / pH-EC | v0.1.1 correction | LoRA plus guarded GGUF/MLX published | LoRA, GGUF/Ollama, MLX | pH/EC risk labels and fertigation blocking |
Model Count¶
- 4 published model families: Agronomist, Tomato, Water/Irrigation, and Actuator Gate.
- 2 active unpublished model families: Sensor Quality and Safety Triage. Nutrient/pH-EC v0.1.1 is published with guarded runtime formats.
- Historical rejected experiments remain local and are not release candidates, including Tomato v0.2.1-v0.2.4, Water v0.1.1-v0.1.7, and rejected Actuator Gate attempts.
- Water v0.1.8 also has published GGUF/Ollama and MLX deployment formats. These are conversions of the canonical LoRA, not separate trained families.
Nutrient / pH-EC v0.1.1¶
The current local candidate uses Qwen/Qwen2.5-0.5B-Instruct with a PEFT LoRA
adapter and explicit derived pH/EC state features. On the independent 140-case
holdout it achieved valid JSON, allowed labels/actions, label F1,
blocked-action F1, and human-review match of 1.0. Rationale wording varies,
so exact object matching is 0.0; this is not a reason to bypass deterministic
validation.
The model is advisory only. It must never dose nutrients, alter pH/EC, change fertigation, or control equipment. Pomona's deterministic rules and human review remain authoritative.
Runtime Formats¶
For accepted candidates, Pomona can prepare PEFT LoRA, GGUF/Ollama, and MLX deployment formats. Nutrient GGUF and MLX builds are prepared by local scripts only for local testing and are not published by this repository. Every conversion requires a fresh runtime evaluation. A conversion is not a new trained model family.
Publication state is tracked separately from technical maturity. A local release candidate is not public until the owner explicitly approves its Hugging Face upload.
See PUBLISHING_SCHEMA.md for the required states and owner-approval workflow.
Safety Rule¶
sensor packet -> specialist model -> schema/label validation
-> deterministic Pomona safety rules -> human approval -> optional automation
See the model registry and LOCAL_MODEL_RUNTIMES.md.