README: add HF download snippet, document 2-column CSV input, swap example paths to released checkpoint

README.md

@@ -116,13 +116,50 @@ python train_ppiDCE.py \
 - `--checkpoint path.pth`: Resume from a saved checkpoint
 - `--suppress_warnings`: Suppress tokenizer truncation warnings
 
+### Quick start: fetch the checkpoint from Hugging Face
+
+The released MED4 checkpoint (`checkpoints/ppiDCE_epoch8.pth`, 12-layer)
+lives on this Hugging Face repo. Pull it without cloning the GitHub mirror:
+
+```python
+from huggingface_hub import hf_hub_download
+
+ckpt_path = hf_hub_download(
+    repo_id="kouroshSA/ppiDCE",
+    filename="checkpoints/ppiDCE_epoch8.pth",
+)
+print(ckpt_path)   # pass this string to --model_path
+```
+
+`inference_ppiDCE.py` takes the checkpoint path as a direct `--model_path`
+argument, so no rename or specific directory layout is required — point
+it straight at the file you just downloaded.
+
+### Input file format
+
+The inference script expects a CSV with two columns of plain amino-acid
+sequences (one protein pair per row — no delimiter tokens, no length
+markers, no chevrons):
+
+```
+seq1,seq2
+MKLR...QSH,MSEDF...VKN
+MQAG...PIA,MTRRL...EEP
+```
+
+A ready-made example is shipped with the repo:
+[`MED4-PPIs-low-confidence_ppiTEPM_prompts.csv`](MED4-PPIs-low-confidence_ppiTEPM_prompts.csv).
+The labeled PRS/RRS reference sets (`MED4_PRS_100.csv`, `MED4_RRS_100.csv`)
+include a third label column, which the inference script ignores — only
+the first two columns are read.
+
 ### Inference
 
 ```bash
 python inference_ppiDCE.py \
-    --model_path out/ppiDCE_epoch8.pth \
+    --model_path checkpoints/ppiDCE_epoch8.pth \
     --model_config facebook/esm1b_t33_650M_UR50S \
-    --input_file test_pairs.csv \
+    --input_file MED4-PPIs-low-confidence_ppiTEPM_prompts.csv \
     --output_file predictions.csv \
     --batch_size 4 \
     --max_length 1024 \