pomona-nutrient-ph-ec-reasoner-v0.1.1-GGUF

Pomona Nutrient / pH-EC Reasoner v0.1.1 GGUF

Experimental runtime artifact, published for local testing. This F16 GGUF is a conversion of the Pomona Nutrient/pH-EC v0.1.1 correction LoRA merged into Qwen/Qwen2.5-0.5B-Instruct.

The source adapter's independent 140-case semantic evaluation passed valid JSON, allowed labels/actions, nutrient label F1, blocked-action F1, and human-review match at 1.0. The conversion was evaluated with the exact training prompt on the same 140-case holdout and is not release-ready. Re-measured on 2026-10-25 (Ollama 0.40.3, temperature 0): valid JSON 1.0000, required fields present 0.9500, allowed-label rate 0.7857, nutrient-label F1 0.5214, blocked-action F1 0.8571, human-review match 0.8571 (evaluation_2026-10-25.json). Normal and missing-critical-data cases remain the main failures. These runtime results do not represent the quality of the source PEFT adapter.

Why these numbers changed. The first measurement (2026-07-19, evaluation.json) reported allowed-label rate 0.8571, nutrient-label F1 0.6690 and required fields 1.0000. Those numbers could not be reproduced: the GGUF tensors are byte-identical, and the water and tomato GGUF evaluations from July and September reproduce exactly on the same Ollama version, so a runtime change is unlikely. The evaluation script was edited shortly after the first run, which is the most likely cause; the earlier version was not kept. evaluation.json is left unchanged as the historical record.

Input format matters. The holdout inputs include a derived_signals block (for example "ph_state": "high") that a rules layer computes. Without it, nutrient-label F1 on the same 140 cases falls to about 0.14–0.19, and the model often lists present fields as missing. Send derived_signals, or better, keep the deterministic route in front of the model.

The Pomona guarded hybrid path applies deterministic rules after model generation. With that guard enabled, the same 140-case holdout reached valid JSON, allowed labels/actions, nutrient label F1, blocked-action F1, and human-review match of 1.0000 in every category (re-confirmed unchanged on 2026-10-25). This measures the complete Pomona deployment path, not standalone GGUF reasoning.

Use only as an advisory component behind Pomona's deterministic nutrient and safety routes. Never use it to directly change fertigation, dose chemicals, control actuators, or diagnose disease.

Local Ollama name prepared: pomona-nutrient-ph-ec:v0.1.1.

Related platform: okyanu/pomona

Usage

Pull directly from Hugging Face with Ollama:

ollama pull hf.co/Okyanus/pomona-nutrient-ph-ec-reasoner-v0.1.1-GGUF

Verified on a clean pull: 994 MB download, ~1.1 GB RAM while loaded (100% GPU offload reported on Apple Silicon), under 2 seconds per inference once loaded. The local Ollama wrapper used for internal testing was:

FROM pomona-nutrient-ph-ec-v0.1.1-f16.gguf
PARAMETER temperature 0
PARAMETER num_predict 220

Use the exact task prompt and keep deterministic validation in front of any fertigation workflow. A prompt or runtime wrapper is not a substitute for the deterministic Pomona safety route. Model-only output is not guaranteed schema-perfect even with format: json — see the evaluation numbers above.

For the guarded path, use scripts/models/guard_nutrient_ph_ec_output.py or the Pomona model-router deterministic route. guarded_evaluation.json records the guarded result; evaluation_2026-10-25.json records the current unguarded model-only result and evaluation.json the original 2026-07-19 run.

Release Status

This package is structurally uploadable as an experimental conversion, but it is not an approved production or release-candidate runtime. The canonical PEFT LoRA remains the reference artifact until this conversion passes an independent full holdout.