QLoRA fine-tune (r=32) of ibm-granite/granite-4.1-3b for German grammar tutoring, quantized to 4-bit (group size 64, affine) for on-device MLX inference. Sibling of granite33-2b-german-tutor-v4-4bit trained on the same corpus; this one trades ~0.4 GB more weight for markedly better error-catching.

Scores

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

suite	items	granite-4.1 v4	granite-3.3 v4	stock 4.1
core grammar (v0)	60	47 (78%)	46 (77%)	31 (52%)
extension (v1ext)	61	46 (75%)	39 (64%)	39 (64%)
holdout (v2)	82	71 (87%)	68 (83%)	41 (50%)
combined	203	164 (81%)	153 (75%)	111 (55%)

Behavior: false-correction rate 9%, miss rate 25% (the 3.3 sibling: 0% / 43% — this model actually attempts corrections instead of defaulting to OK). Naturalness: modal particles 12.0 per 100 tokens, follow-up rate 0.82.

Paired vs the 3.3 sibling on identical items: 26 gained / 15 lost (exact McNemar p = 0.12).

Notes

~100k vocabulary tokenizes German at roughly half the tokens of the 3.3 sibling's 49k BPE, largely offsetting the larger parameter count in per-sentence latency.
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), LanguageTool + spaCy validated.

Output contract

OK for correct sentences, otherwise:

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

kessenma/granite41-3b-german-tutor-v4-4bit

Granite 4.1 3B German Tutor v4 (MLX 4-bit)

Scores

Notes

Output contract

kessenma/
granite41-3b-german-tutor-v4-4bit