synoema-coder-3b-v3

Language 0.1.0-beta.1

General Synoema code generation from natural language prompts.

Trained on Synoema — a formally verified functional language (GBNF + Hindley-Milner + contracts, prompt to native/WASM/IoT with no human review).

Evaluation

Metric	Value
run_pass	70.2%
compile_pass	67.3%
Eval set	104 examples
Method	greedy (do_sample=False)
Language version	0.1.0-beta.1

Quickstart

wget https://huggingface.co/delimitter/synoema-coder-3b-v3/resolve/main/synoema-coder-3b-v3-q4km.gguf
wget https://huggingface.co/delimitter/synoema-coder-3b-v3/resolve/main/Makefile
make pull
make run

make pull sets the system prompt automatically — required for correct behavior.

Target	Action
make pull	Create Ollama model + system prompt
make run	Interactive chat
make clean	Remove from Ollama

Example prompts

Write a recursive fibonacci function
Define a map function for lists
Write quicksort with pattern matching
Create a function with requires/ensures contracts

Model details

Field	Value
Base model	Qwen/Qwen2.5-Coder-3B-Instruct
Fine-tuning	QLoRA SFT (LoRA r=32)
Size	1.9 GB Q4_K_M
Language version	0.1.0-beta.1

Links

• synoema.tech · docs · iot · all models

GGUF Downloads (Ollama / llama.cpp)

File	Size	Recommended for
synoema-coder-3b-v3-Q4_K_M.gguf	1.8 GB	Most users (CPU/GPU)
synoema-coder-3b-v3-Q8_0.gguf	3.1 GB	High accuracy

Ollama Quickstart

ollama run hf.co/delimitter/synoema-coder-3b-v3:Q4_K_M