Exercises and answers come from a checked-in
language pack — no model is involved, so the key is correct by
construction. Available for French, German and Romanian.
Written by the model, so the key prints
Model answer — one acceptable answer, not
the only one. The story is stored as written, so a reprint
is the same paper.
Choose skills and press Preview.
What you see here is exactly what prints.
Tutor Sessions
⚙️ Tutor LLM Parameters
Temperature0.6
Controls randomness. 0 = deterministic, 1 = creative. 0.6 is good for natural conversation.
Top K40
Limits token selection to K most likely tokens. Lower = more focused, higher = more diverse.
Top P0.9
Nucleus sampling — considers tokens whose cumulative probability reaches P.
Min P0.0
Minimum token probability threshold. Filters out unlikely tokens. 0 = disabled.
Repeat Penalty1.1
Penalizes token repetition. 1.0 = no penalty, higher = less repetition. 1.1 is a good default for text.
Frequency Penalty0.0
Reduces repeated tokens proportional to frequency. 0 = disabled.
Presence Penalty0.0
Encourages new topics. Penalizes any token that appeared. 0 = disabled.
Max Tokens (num_predict)4096
Maximum tokens per response. 4096 is sufficient for tutor conversations. Increase for longer lessons.
Context Length (num_ctx)32768
Context window size in tokens. Larger = more conversation memory but uses more VRAM.
Seed0
Random seed for reproducibility. 0 = random each time.
📚 Tutor Agent
Max Iterations (max_iter)15
Maximum reasoning/tool-call iterations per task. Too low = incomplete answers. 10–20 recommended for tutor.
Planning Max Attempts3
Maximum planning attempts before proceeding. Higher = more retries if planning fails. 1–5 recommended.
Request Timeout (sec)300
Maximum seconds the agent may run before timing out. Increase for long lessons or complex quizzes. 60–900 recommended.