{ "experiment": "Gemma 4 V4 input-variation SAE (bnb 4-bit)", "caveat_quantization": "4-bit quantized inference; results characterize quantized checkpoint.", "prompt_equivalence": "DIAGNOSTIC ONLY: inputs DIFFER across neutral / V0 / V3 templates. This SAE is an input-classifier baseline, NOT a behavior detector. Cohen's d here captures input-template separability, not generation-process separability.", "model": { "key": "gemma4_e2b", "hf_name": "google/gemma-4-E2B", "d_in": 1536, "layer": 10, "quantization": "bnb 4-bit nf4 double_quant fp16_compute" }, "sae": { "architecture": "topk", "k": 32, "d_sae": 6144, "metrics": { "mse_loss": 0.007008623331785202, "explained_variance": 0.9957067616251516, "l0": 32.0, "alive_features": 559, "train_time_s": 108.4, "num_activations": 3619, "final_epoch_loss": 0.00756 }, "checkpoint": "gemma4_e2b_v4input_topk_k32_L10_bnb4bit.pt" }, "input_split": { "n_neutral": 50, "n_v0": 50, "n_v3": 50, "neutral_vs_v0_block": { "label": "input_neutral_vs_v0", "d_max": 16.963083267211914, "d_mean": 0.011611283756792545, "d_above_1": 18, "d_above_2": 8, "top10_ids": [ 4775, 1592, 1349, 4973, 2951, 849, 5594, 5259, 4368, 5822 ], "top10_d": [ 16.963083267211914, 8.70340633392334, 3.4786453247070312, 3.3507778644561768, 3.049485206604004, 2.4634058475494385, 2.374709367752075, 2.126396656036377, 1.9656071662902832, 1.7550815343856812 ] }, "neutral_vs_v3_block": { "label": "input_neutral_vs_v3", "d_max": 8.594282150268555, "d_mean": 0.010855759494006634, "d_above_1": 20, "d_above_2": 10, "top10_ids": [ 1592, 5594, 3907, 5560, 4020, 4912, 849, 995, 1349, 3968 ], "top10_d": [ 8.594282150268555, 7.050577163696289, 5.138309478759766, 3.4406070709228516, 3.346367835998535, 2.9664735794067383, 2.887408494949341, 2.620266914367676, 2.325357675552368, 2.1497621536254883 ] } }, "timestamp": "2026-04-21T06:46:21.923664+00:00" }