| { | |
| "interpretability_model": "Qwen/Qwen2.5-1.5B-Instruct", | |
| "n_layers": 28, | |
| "best_probe_layer": 23, | |
| "best_probe_accuracy": 0.6514492753623189, | |
| "note": "This model is a dedicated white-box interpretability testbed, distinct from the 16-model black-box leaderboard." | |
| } |