TestWiz
- #E-Learning
The task was to develop a solution for answering multiple-choice questions on MCAT and SAT tests using the OpenAI GPT-3/4 Davinci model. The goal was to create custom fine-tuned models for both tests and achieve a high accuracy rate for answering multiple-choice questions.
- Natural Language Processing
Impact
The platform allows:
- Primary challenge: Ensuring accuracy of fine-tuned models, overcome by dataset expansion and specific test format optimization.
- Availability of quality datasets posed another challenge, requiring extensive collection and curation efforts.
Services we provided
Promising results achieved through testing custom fine-tuned models
Accuracy rate of 0.62 obtained, marking a significant improvement
Our solution offers time savings, increased productivity, and enhanced test question analysis and evaluation
Tech Stack
GPT 3.5
GPT 4
Challenges and Solutions
🧐 Challenges
- Primary challenge: Ensuring accuracy of fine-tuned models, overcome by dataset expansion and specific test format optimization.
- Availability of quality datasets posed another challenge, requiring extensive collection and curation efforts.
💡 Solutions
- Initially challenging fine-tuning of the OpenAI GPT-3/4 Davinci model on MCAT and SAT datasets.
- Dataset expanded in quantity and quality by gathering more test questions and ensuring accuracy.
- -Separate models trained for MCAT and SAT tests, with custom evaluation emphasizing answer accuracy.
User flow
1. The user selects either MCAT or SAT and submits questions.
2. These questions are processed using the fine-tuned Davinci model.
3. The participant receives accurate answers and analysis.
4. Outcomes help refine the study/test strategy.