README.md

metadata

license: mit
task_categories:
  - text-classification
tags:
  - medical
  - biology

Data Selection & Splitting

• Source: HRCS 2014, 2018, and 2022 direct award datasets.
• Quality Filtering:
  • Only human-coded abstracts were included.
  • Records with abstracts shorter than 75 characters were removed during preprocessing to ensure the model had sufficient text to learn from.
• Train/Test Split: The Test Set was isolated using only 2022 data to provide a modern performance benchmark.

---

Training Data Deduplication

To prevent the model from over-fitting on near-identical entries, a robust deduplication pipeline was implemented:

1. Vectorization: Character-level TF-IDF vectors were generated from training titles using word-boundary character n-grams (length 3–5).
2. Similarity Analysis: Near-duplicate titles were identified using a Cosine Similarity threshold of more than or equal to 0.85.
3. Clustering: Records exceeding this threshold were grouped using a connected-components graph algorithm.
4. Selection: Only the first occurrence in file order from each group was retained in the training pool.

---

Leakage Prevention (Train vs. Test)

To ensure the test set provides a truly unseen and honest evaluation, the following steps were taken:

• Shared Feature Space: The TF-IDF vectorizer was fit on the combined set of training and test titles.
• Cross-Set Comparison: Any training record with a Cosine Similarity threshold of more than or equal to 0.85 to any record in the test set was permanently removed from the training pool.
• Test Set Integrity: The test set itself was deduplicated using exact title matching only (no fuzzy matching applied).

Limitation: Short, highly generic grant titles (e.g., "Studentship") may have been deduplicated in the training set due to the similarity threshold.