Synoema IoT Rules model — fine-tuned for generating Synoema IoT automation rules that compile and run on bare-metal MCUs, ESP32, STM32, and Linux edge devices.

Synoema is a formally verified, BPE-aligned functional language for LLM-generated software. GBNF grammar + Hindley-Milner types + contracts eliminate the verification gap — prompt to native / WASM / IoT with no human review.

🌐 synoema.tech · 📦 GitHub · 📖 Language Reference

Model Description

Smallest IoT model (0.5B parameters). Best compile rate. Ideal for resource-constrained MCUs.

This model generates Synoema IoT rules — compact functional programs that:

Compile to WASM artifacts (mean size ~82–292 bytes)
Run on Tier 0 bare-MCU (wasm3), Tier 1 ESP32/STM32, and Tier 2 Linux edge
Express automation logic with formal requires/ensures contracts
Support verticals: Home, Industrial, Wearable, Automotive, Agriculture

Evaluation Results

Split	compile_pass	Total	Pass Rate
Val set	101	104	97.1%
Test set (unseen)	124	127	97.6%

Threshold: ≥80/104 (≥76.9%) — PASSED ✅

GGUF (llama.cpp / Ollama)

File: synoema-iot-nano-0.5b-v1-q3km.gguf (339 MB, Q3_K_M)

bash

# Ollama
ollama run delimitter/synoema-iot-nano-0.5b-v1

# llama.cpp
./llama-cli -m synoema-iot-nano-0.5b-v1-q3km.gguf \
  --system-prompt "You are a Synoema IoT rules expert. Generate only valid Synoema IoT rule code." \
  -p "activate the pump if humidity falls under 35"

Usage (Python / transformers)

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base = "Qwen/Qwen2.5-Coder-0.5B-Instruct"
adapter = "delimitter/synoema-iot-nano-0.5b-v1"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

SYS = "You are a Synoema IoT rules expert. Generate only valid Synoema rule code. No explanation."
prompt = tok.apply_chat_template([
    {"role": "system", "content": SYS},
    {"role": "user",   "content": "activate the fan if temperature exceeds 30"},
], tokenize=False, add_generation_prompt=True)

inp = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inp, max_new_tokens=128, do_sample=False)
print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))

Training Details

Parameter	Value
Language version	`0.1.0-beta.1`
Training corpus	`corpus_iot_rules_train.jsonl` (946 examples, Wave-2)
Method	QLoRA (LoRA r=16, α=32)
Base model	`Qwen/Qwen2.5-Coder-0.5B-Instruct`
Epochs	3
Batch size	4 (eff.)

Synoema IoT Rule Example

synoema

rule_fan_control temp prev =
  ? temp > 30 -> 1 : ? temp < 25 -> 0 : prev

main = rule_fan_control 35 0

License

Apache 2.0. Base model license applies separately. See synoema.tech for full terms.