granite33-2b-german-tutor-v4-4bit

Granite 3.3 2B German Tutor v4 (MLX 4-bit)

QLoRA fine-tune (r=32) of ibm-granite/granite-3.3-2b-instruct for German grammar tutoring, quantized to 4-bit (group size 64, affine) for on-device MLX inference. Built as the small-memory tier of an iOS flashcard app: at ~1.4 GB it runs on devices that cannot hold a larger tutor.

Supersedes the unreleased r32 build trained on the earlier corpus.

Scores

App-guard convention throughout (a reply counts only if the app's parser would show it).

Paired vs the previous tune: 35 items gained, 13 lost, exact McNemar p = 0.002.

Behavioral profile: 0% false corrections (it never marks correct German wrong; unchanged from the previous tune) and a 43% miss rate on erroneous sentences (improved from 59%, but still the model's main limitation — when unsure it says OK). Conversational naturalness is strong: modal particles 11.9 per 100 tokens, follow-up question rate 0.92.

Notes

• The tokenizer is a 49k byte-level BPE: German text needs roughly 1.9× the tokens of a SentencePiece model of similar size, so generation is slower per word than the Gemma tutors.
• Trained on corpus v4: 42,841 rows, 15 grammar phenomena at 70% fix / 30% ok per phenomenon, teacher-generated (gemma-4-31B bulk + Claude Sonnet judgment phenomena), all rows validated with LanguageTool + spaCy.

Output contract

OK for correct sentences, otherwise:

FIX: <corrected sentence>
WHY: <one-line explanation>
HINT: <question that nudges the learner>