EuroSafeGuard v1: probe, 5 input-classifier seeds, exchange classifier, thresholds, loader, model card
b8ee2d2 verified | { | |
| "kind": "mlp", | |
| "hidden": 256, | |
| "activation": "gelu", | |
| "base_model": "meta-llama/Llama-3.1-8B-Instruct", | |
| "layers": [ | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9, | |
| 10, | |
| 11, | |
| 12, | |
| 13, | |
| 14, | |
| 15, | |
| 16, | |
| 17, | |
| 18, | |
| 19, | |
| 20, | |
| 21, | |
| 22, | |
| 23, | |
| 24, | |
| 25, | |
| 26, | |
| 27, | |
| 28, | |
| 29, | |
| 30, | |
| 31, | |
| 32 | |
| ], | |
| "hidden_size": 4096, | |
| "feature_dim": 131072, | |
| "features": "concatenated hidden_states[l] for l in layers (outputs of decoder blocks 1-32), bf16", | |
| "scored_positions": "prompt only: the chat-formatted user turn up to and including the assistant header (no thinking)", | |
| "max_len": 2048, | |
| "inference": "EMA of per-token logits, alpha = 2/(M+1), exchange score = max over positions", | |
| "M": 16, | |
| "val_auroc": 0.9442754882489239 | |
| } |