Fine-tuned from openbmb/MiniCPM4-8B for exam-question generation in PaperProf, an AI study buddy that turns course PDFs into interactive quiz sessions.

Training

Method: QLoRA (4-bit NF4, r=16, alpha=32, all-linear targets), merged to bf16
Data: ~3500 multi-task pairs in PaperProf's three production formats: open question generation (SQuAD), MCQ with distractors and per-option explanations (SciQ), and structured answer evaluation (SQuAD-derived), so the model is optimized for the exact tasks it serves.
Epochs: 1, lr 2e-4 cosine, bf16 compute

Usage

Drop-in replacement for the base model:

python

from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("build-small-hackathon/MiniCPM4-8B-PaperProf", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("build-small-hackathon/MiniCPM4-8B-PaperProf", trust_remote_code=True, torch_dtype="bfloat16")

Built for the Build Small Hackathon, June 2026, by Team PaperProf (EPITA).