DoctorPuppy/studyapp-writer

StudyApp card writer

The card writer that the StudyApp desktop app downloads. It's Gemma 4 E2B (instruction-tuned), trained with a LoRA adapter on StudyApp's own tasks, merged, and quantized to Q4_K_M for llama.cpp. It runs locally, entirely offline.

The tasks it's trained on, using the app's own prompts and output formats:

• flashcards for each page of a student's notes
• facts, scenario questions and practice-test questions
• grading a typed answer, and checking the key points of a written one
• chat that answers only from the student's notes, and answers NOT_FOUND when the notes don't have the answer

StudyApp runs it with the multi-token-prediction head from ggml-org/gemma-4-E2B-it-GGUF (mtp-gemma-4-E2B-it-Q8_0.gguf) for faster generation.

Training data

There are 618 examples. Each one was written for excerpts of openly licensed textbooks and passed the same checks the app runs on its own output. 569 were used for training and 49 were held out to measure it.

The excerpts come from these OpenStax textbooks (openstax.org), which are licensed under CC BY 4.0. They're free to read at openstax.org.

• American Government 2e
• Anatomy and Physiology
• Biology
• Chemistry
• College Physics
• Introduction to Sociology 2e
• Introductory Statistics
• Precalculus
• Principles of Economics 2e
• Psychology

Release test

This writer was compared with the original Gemma 4 E2B on two textbooks it never trained on: OpenStax Astronomy and Introduction to Intellectual Property.

	Original	This writer
Flashcards: answers that broke the app's format or limits	0 of 46	0 of 46
Flashcards: cards the app's checks flagged	0 of 136	3 of 145
Flashcards: blind read, which writer's cards were better	6	22 (18 the same)
Chat: questions the notes answer, answered cleanly	12 of 12	11 of 12
Chat: questions the notes can't answer, declined	9 of 11	10 of 11
Practice-test questions: problems found in 3 page batches	3	8

License

The model weights are licensed under Apache-2.0, the same as Gemma 4.