Evora on-device text model (Qwen3-1.7B, v3)

The first release of the text model as one English base with one LoRA adapter per further language, for the Evora apps. The apps load the base and apply the adapter of the user's language at run time (llama.cpp), so a language is a download of about 25 MB instead of another model.

Same tasks, prompt format and JSON answers as v2 — see its card for the prompt contract and samples. With llama.cpp:

llama-server -m evora-qwen3-1.7b-en.Q4_K_M.gguf --lora evora-qwen3-1.7b-de.lora.f16.gguf --jinja --temp 0

Training

• Base: Qwen/Qwen3-1.7B, LoRA (r 32, alpha 64) on q, k, v, o, three epochs over 2,797 English rows of the five tasks (answers by gpt-5.4 to Evora's production prompts, synthetic users), merged, converted and quantized with llama.cpp. Trained in AI Studio on a rented GPU.
• German adapter: LoRA on this base, two epochs over 3,505 German rows.

Evaluation

Every held-out test row (139 English, 148 German; synthetic users), answered without a JSON grammar and checked against each task's exact schema -- every field present, no other field -- next to v2.1, the model it replaces:

The German difference is well outside the noise (0.079 ± 0.004 per prompt). Measured on the unquantized weights in AI Studio; the phones run the Q4_K_M base with the f16 adapter.