Add MED4 epoch-4 checkpoint used for low-confidence Y2H screening, with sidecar note

checkpoints/ppiBTPE_epoch_4.md

@@ -0,0 +1,34 @@
+# ppiBTPE_epoch_4.pth
+
+**Checkpoint used for screening low-confidence Y2H pairs in the *Prochlorococcus marinus* MED4 interactome.**
+
+## Provenance
+
+| | |
+|---|---|
+| Model | ppiBTEP (Siamese / twin-branch ESM-1b-inspired transformer, trained from scratch) |
+| Architecture | 12 transformer layers |
+| Epoch | 4 |
+| File size | ~913 MB |
+| Training run | `out_june17_MED4_l12` (June 17, 2025) |
+| Training set | `train_MED4_ppiBTEPM-pseudo_Int_combo1-2-3.csv` (≈13,008 pairs, pre-clean — see note below) |
+| Validation set | `val_MED4_100_Y2H-RND_ppiBRTPM.csv` |
+
+## Intended use
+
+Inference / screening of candidate MED4 protein–protein interactions that
+were originally flagged as **low-confidence Y2H hits**. The model is run on
+each candidate pair and its softmax probability is used (in concert with the
+other tri-model components, ppiDCE and ppiGPLM) to retain or discard the pair.
+
+## Notes
+
+- This checkpoint was produced **before** the PRS/RRS de-overlapping pass on
+  `train.csv` (see [`MED4-PPIs-low-confidence_ppiTEPM_prompts.csv`](../MED4-PPIs-low-confidence_ppiTEPM_prompts.csv) and the cleaned
+  `train.clean.csv` / `train.clean2x.csv` companions). Approximately 608 of
+  the 13,008 training rows (4.67 %) overlap with the PRS+RRS evaluation pairs
+  in either orientation. Treat metrics on those pairs accordingly.
+- Loading: use `train_ppiBTPE3b.py` / `inference_ppiBTPE_2GPU.py` from the
+  parent repo with `--num_layers 12` and `--model_config
+  facebook/esm1b_t33_650M_UR50S` (config-only; weights are loaded from this
+  checkpoint, not from the HF ESM-1b release).