Update README.md

README.md

@@ -1,25 +1,33 @@
 ---
 license: mit
 ---
+## Data Selection & Splitting
 
-1. Data Selection & Splitting  
-Source: HRCS 2014, 2018, and 2022 direct award datasets.
+* **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.
 
-Quality Filter: Records with abstracts shorter than 75 characters were removed during preprocessing prior to the Python deduplication and filtering steps to ensure the model had sufficient text to learn from. Only human coded abstracts were included.
+---
+
+## 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 $\ge 0.85$**.
+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.
+
+---
 
-Train x Test split: The Test Set was isolated using only 2022 data to provide a modern performance benchmark.
+## Leakage Prevention (Train vs. Test)
 
-2. Training Data Deduplication  
-Character-level TF-IDF vectors were generated from training titles using word-boundary character n-grams of length 3–5.  
-Cosine similarity ≥ 0.85 was used to identify near-duplicate titles.  
-Records exceeding this threshold were grouped using a connected-components graph algorithm.  
-Only the first occurrence in file order from each group was retained in the training pool.
+To ensure the test set provides a truly unseen and honest evaluation, the following steps were taken:
 
-3. Leakage Prevention (Train vs. Test)  
-To ensure the test set provides a truly unseen, honest evaluation of the model, the training pool was compared against the test data using the same cleaning procedure, TF-IDF character n-gram representation, and cosine similarity threshold (≥ 0.85).  
-For this step, the TF-IDF vectorizer was fit on the combined set of training and test titles to ensure a shared feature space.  
-Any training record with a cosine similarity ≥ 85% to any record in the test set was permanently removed from the training data.  
-The test set itself was deduplicated using exact title matching only (no fuzzy matching was applied).
+* **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 $\ge 85\%$** 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).
 
-Limitations:  
-Short, highly generic grant titles (e.g., "Studentship") may have been deduplicated in the training set due to the similarity threshold.
+> [!IMPORTANT]
+> **Limitation:** Short, highly generic grant titles (e.g., *"Studentship"*) may have been deduplicated in the training set due to the similarity threshold.