Downloads · 30 days
0
tenroman/qwen3-vl-8b-flashcard-qlora
qwen3-vl-8b-flashcard-qlora is a machine learning model from tenroman. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
This project explores whether parameter-efficient fine-tuning can teach a local language model to convert unstructured educational passages into concise, grounded JSON flashcards.
Downloads · 30 days
0
Access
Public
Updated Jul 19, 2026
Repo size
38.8 MB
Likes
1
Public
Click a slice to open those files.
.safetensors38.8 MB · 100%
From the Hugging Face model README
This project explores whether parameter-efficient fine-tuning can teach a local language model to convert unstructured educational passages into concise, grounded JSON flashcards.
I fine-tuned a 4-bit Qwen3-VL-8B-Instruct model using QLoRA and MLX-LM on a 24 GB Apple-silicon MacBook. The training dataset was built from 50 hand-authored seed examples, expanded into 500 synthetic examples, structurally validated, deduplicated, and divided into training and validation sets. A separate set of 20 manually curated examples was reserved exclusively for final evaluation.
The untouched base model already generated valid flashcards but tended to create too many cards by treating nearly every clause as an independent fact. The fine-tuned adapter learned a more selective behavior, reducing total generated cards from 57 to 41 against a 35-card reference set. Exact card-count alignment improved from 5 of 20 examples to 15 of 20, while all evaluated outputs remained valid JSON and preserved the supplied source text exactly.
The experiment also exposed limitations: the adapted model occasionally omitted useful secondary facts, retained sensitivity to deliberate distractors, and reproduced recurring phrasing patterns from the synthetic dataset. These results demonstrate both the effectiveness of task-specific QLoRA and the importance of dataset diversity and evaluation design.