Upload README.md with huggingface_hub

README.md

@@ -73,19 +73,53 @@ shift. Every fitted statistic — vocabulary, quantile edges, scalers — is fit
 on training patients only. 152 automated checks cover this, including a grep
 test that no module outside the time utility parses a timestamp.
 
-## Reproducing
-
-The data is not included here. Rebuilding `predictions.csv` from the files in
-this repository, without any training:
-
-1. Clone the code and install it:
-   `git clone https://github.com/Sallamsaka/M31-Coding-Test`, then
-   `pip install -r requirements.txt` inside it.
-2. Copy the provided data into the repository root: the folders `train_val/` and
-   `test/`, and the files `patient_splits.csv`, `target_conditions.csv` and
-   `test_anchors.csv`.
-3. Run `python -m src.reproduce --from-hub sallamsaka/M31-Coding-Test`. It downloads
-   these model files, predicts every test patient, and checks the result against
-   the submitted `predictions.csv`.
+## Reproducing the predictions
+
+The files on this page are enough to rebuild the submitted `predictions.csv`
+exactly, without any training. The patient data is not included here: you need
+the dataset provided with the M31 take-home.
+
+**You need** Python 3.13, git, and that dataset.
+
+1. Get the code and install its dependencies:
+
+   ```
+   git clone https://github.com/Sallamsaka/M31-Coding-Test
+   cd M31-Coding-Test
+   pip install -r requirements.txt
+   ```
+
+2. Copy the provided data into that folder, so that it looks like this:
+
+   ```
+   M31-Coding-Test/
+     train_val/              provided CSV tables, train and validation patients
+     test/                   provided CSV tables, test patients
+     patient_splits.csv
+     target_conditions.csv
+     test_anchors.csv
+     src/, outputs/, ...     already there from the clone
+   ```
+
+3. Run:
+
+   ```
+   python -m src.reproduce --from-hub sallamsaka/M31-Coding-Test
+   ```
+
+   This downloads the model files from this page, builds each test patient's
+   features and event sequence, runs logistic regression, the boosted trees and
+   the three transformer seeds, averages and calibrates them, and writes
+   `outputs/predictions_reproduced.csv` (358 patients, 40 conditions).
+
+4. Check the last line it prints. The clone already contains the submitted
+   `outputs/predictions.csv`, and the script compares against it:
+
+   ```
+   max |reproduced - outputs/predictions.csv| = 7.40e-08  (MATCH)
+   ```
+
+   Anything below 1e-5 prints `MATCH`; the remaining difference is floating-point
+   rounding.
 
 Retraining everything from scratch is described in the GitHub README.