---

license: apache-2.0

base_model: Qwen/Qwen3-4B-Instruct-2507

language:

• de
• en

pipeline_tag: text-generation

tags:

• gguf
• qwen3
• german
• language-learning
• on-device

---

Laura — German tutor in English (Qwen3 4B Instruct 2507, GGUF)

Laura is the tutor in an offline German course app for adult beginners

(A1–A2). This model talks to the learner in English and teaches German with

German examples. It runs on the phone, CPU only.

This is a modified version of Qwen/Qwen3-4B-Instruct-2507.

Changes made: a LoRA adapter (rank 16, all linear layers, 2 epochs) was trained

on ~2,500–3,300 short tutoring conversations in English, merged into the weights,

and the result converted to GGUF and quantised to Q4_K_M with llama.cpp.

Distributed under the base model's Apache-2.0 licence.

How the app uses it

The app does not trust the model to judge German on its own. It checks the

learner's sentence with its own dictionary first and adds the verdict to the

message — (Check: ✗ → „corrected sentence“ · reason) or (Check: ✓ …) — plus

dictionary meanings and the chapter's rule card. The model was trained with

exactly those notes and explains them; without them it is a much weaker teacher.

Evaluation

Round 5 (2 Oct 2026): Qwen3-4B-Instruct-2507 with a LoRA on 3,974 English tutor conversations (2 epochs), then a DPO step on 278 preference pairs (the earlier model's own replies that failed the app's checks, against the correct ones). Blind-graded through the full app on 40 fresh held-out conversations (154 replies): 18.2% of replies with a mistake, against 22.7% for round 3 on the same set; mistakes in the model's own text 9 against 16.

Files

• laura-en-qwen3-4b-Q4_K_M.gguf — Q4_K_M, 2.50 GB