{ "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": 17, "quantization": "bnb 4-bit nf4 double_quant fp16_compute" }, "sae": { "architecture": "topk", "k": 32, "d_sae": 6144, "metrics": { "mse_loss": 0.02122976817190647, "explained_variance": 0.9926107570635635, "l0": 32.0, "alive_features": 241, "train_time_s": 110.0, "num_activations": 3619, "final_epoch_loss": 0.021807 }, "checkpoint": "gemma4_e2b_v4input_topk_k32_L17_bnb4bit.pt" }, "input_split": { "n_neutral": 50, "n_v0": 50, "n_v3": 50, "neutral_vs_v0_block": { "label": "input_neutral_vs_v0", "d_max": 7.808701515197754, "d_mean": 0.0064131771214306355, "d_above_1": 11, "d_above_2": 1, "top10_ids": [ 2196, 477, 2905, 4422, 414, 1729, 2933, 2682, 4232, 1926 ], "top10_d": [ 7.808701515197754, 1.8714970350265503, 1.7008979320526123, 1.686379313468933, 1.5654693841934204, 1.2770538330078125, 1.2198415994644165, 1.1726460456848145, 1.1275814771652222, 1.085732102394104 ] }, "neutral_vs_v3_block": { "label": "input_neutral_vs_v3", "d_max": 9.32315731048584, "d_mean": 0.010321254841983318, "d_above_1": 14, "d_above_2": 6, "top10_ids": [ 2196, 3340, 4259, 519, 4232, 477, 3928, 1926, 561, 2905 ], "top10_d": [ 9.32315731048584, 4.7347235679626465, 2.557842254638672, 2.2542951107025146, 2.1295456886291504, 2.0010032653808594, 1.7745753526687622, 1.6784650087356567, 1.6581165790557861, 1.6453344821929932 ] } }, "timestamp": "2026-04-20T08:26:20.056190+00:00" }