RecLetter Strength BERT Tiny
This repository contains the fine-tuned recommendation-strength classifier used in the ISOM5240 project, AI-Assisted Recommendation Letter Drafting System for University Faculty.
Model
- Base model:
prajjwal1/bert-tiny
- Task: text classification
- Labels:
Moderate Recommendation, Strong Recommendation, Exceptional Recommendation
- Intended use: classify transcript-like student evidence into a recommendation-strength signal for a faculty-reviewed recommendation letter draft.
Training Data
The training data is synthetic and privacy-safe. It is generated from structured student profiles rather than real transcripts or real recommendation letters.
Generation logic:
- Moderate profiles use lower GPA bands, mid-range class percentiles, steady coursework language, and basic project evidence.
- Strong profiles use higher GPA bands, top-quartile class percentiles, stronger analytical or communication notes, and developed project evidence.
- Exceptional profiles use top GPA bands, top-decile class percentiles, original research or leadership notes, and advanced project evidence.
Evaluation Note
The model achieved very high held-out accuracy on the synthetic test set. This mainly shows that the model learned the designed recommendation-strength rules. It should not be interpreted as real-world generalization accuracy.
Limitations
The model should be used only as a drafting aid. A production version would require consented student profiles, faculty-reviewed labels, privacy controls, and human approval before any recommendation letter is submitted.
