Add library requirements and UNEXPECTED/MISSING key explanations

README.md

@@ -122,8 +122,31 @@ highlight_indices = [i for i, score in enumerate(scores) if score >= 0.9]
 ### Installation
 
 ```bash
-pip install torch transformers
+pip install "transformers>=4.41" "huggingface_hub>=0.27" torch
 ```
+> **Important:** Requires `huggingface_hub >= 0.27` to avoid deprecated API errors.
+
+### Requirements
+
+| Library | Minimum Version | Notes |
+|---------|----------------|-------|
+| `transformers` | **>= 4.41** | XLMRobertaModel support |
+| `huggingface_hub` | **>= 0.27** | No deprecated `use_auth_token` |
+| `torch` | **>= 2.0** | CUDA support |
+
+```bash
+pip install "transformers>=4.41" "huggingface_hub>=0.27" torch
+```
+
+### Loading Notes
+
+**UNEXPECTED / MISSING keys on load:** When loading `AutoModel.from_pretrained("BAAI/bge-reranker-v2-m3")`, you will see warnings about:
+
+- **UNEXPECTED** keys (`classifier.out_proj.weight`, `classifier.out_proj.bias`, `classifier.dense.weight`, `classifier.dense.bias`): These are the **reranker's classification head** weights. Since we load with `AutoModel` (encoder-only, no classification head), these extra keys are not consumed. This is **completely normal** — we only need the encoder backbone. Our separate `PruningHead` MLP replaces the original classification head.
+
+- **MISSING** keys (`pooler.dense.weight`, `pooler.dense.bias`): The `AutoModel` class expects a pooler layer, but the BGE-reranker checkpoint was trained without one. These get **randomly initialized** but are **never used** in our pipeline — we use `last_hidden_state` directly, not the pooled output. Safe to ignore.
+
+
 
 ### Quick Start