QuantOsimo Maths Marking Model v1
Model Description
A fine-tuned model for marking GCSE, AS-Level and A-Level Mathematics answers against official mark schemes. Given a question, student answer and mark scheme it awards marks and provides examiner-style feedback.
Intended Use
- Automated marking of student answers in EdTech platforms
- Research into automated assessment
- Educational AI development
Training Data
The model was trained on structured data extracted from GCSE (Foundation and Higher), AS-Level and A-Level Mathematics mark schemes covering the years 2018-2025 (including 2023-2024, which the evaluation below draws on) from AQA, Edexcel and OCR: 5,421 examples from 587 papers. The training data is not publicly released due to exam board copyright. A sample of 100 examples demonstrating the data format is available in the repository.
Evaluation Results
Evaluated on 50 questions from 2023 and 2024 exam papers. These papers were part of the training data (see Notes), so these results measure performance on seen material, not generalisation to unseen papers.
Overall Performance
- Attempted correctly: 44/50 (88%)
- Wrong or unreadable input: 6/50 (12%)
By Exam Board
| Board | Score |
|---|---|
| AQA | 88% (15/17) |
| Edexcel | 100% (17/17) |
| OCR | 75% (12/16) |
By Level
| Level | Score |
|---|---|
| GCSE | 91% (31/34) |
| AS-Level | 100% (10/10) |
| A-Level | 50% (3/6) |
Notes
- 5 of the 6 wrong answers were questions with garbled PDF extraction (unreadable input text)
- The A-Level score is based on only 6 questions, too few to be conclusive
- Evaluation used conservative scoring: any response with garbled or unreadable input was scored 0 regardless of model quality
- All evaluation questions were from 2023-2024 papers. v1 was trained on every year from 2018 to 2025, so 2023 and 2024 were not held out: these results likely overstate performance on unseen papers
Evaluation Dataset
50 questions from:
- AQA, Edexcel, OCR
- GCSE Foundation and Higher, AS-Level, A-Level
- Mathematics 2023-2024
- 2-5 marks per question
- No diagram questions
Limitations
- Only trained on Mathematics
- UK curriculum only
- May not generalise to other curricula
How to Use
[To be filled after training]
Citation
Osagiede, S. (2026). QuantOsimo Maths Marking Model v1. https://huggingface.co/Osam-2025/quantosimo-maths-marking-v1
@misc{osagiede2026quantosimomarking,
author = {Osagiede, Samuel},
title = {QuantOsimo Maths Marking Model v1},
year = {2026},
url = {https://huggingface.co/Osam-2025/quantosimo-maths-marking-v1}
}