Evora on-device text model (Qwen3-1.7B, v1)
The on-device fallback for four Evora text tasks: meal_plan.stub, meal_plan.details,
coaching.rate_item, meal.identify.text.
| File | Size |
|---|---|
evora-qwen3-1.7b-v1.Q4_K_M.gguf |
1.0 GB |
Prompt
System prompt (the model was trained on exactly this; the task and language vary):
You are Evora's on-device model. Task: {task}. Language: de (German). Reply with the JSON object only.
User message: the same JSON payload the Evora API sends to its server model for that task. The answer is one compact JSON object in the Evora API's production schema.
Qwen3 must be prompted with thinking off (enable_thinking: false in the chat template,
--jinja in llama.cpp); apply a JSON grammar only after that, or the grammar forces the
model's reasoning into a string.
Training
Distilled from gpt-5.4 answering Evora's production prompts, in AI Studio. LoRA on the language model, merged, converted with llama.cpp. English and German.
Data: 5,168 rows (inputs shaped like the app's real payloads; item ratings on Open Food Facts products, ODbL). Held-out: 100% valid JSON in the production schema; item grades within one step of gpt-5.4 91% of the time.