{ "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": 25, "quantization": "bnb 4-bit nf4 double_quant fp16_compute" }, "sae": { "architecture": "topk", "k": 32, "d_sae": 6144, "metrics": { "mse_loss": 0.014618407003581524, "explained_variance": 0.9957529839731779, "l0": 32.0, "alive_features": 394, "train_time_s": 109.1, "num_activations": 3619, "final_epoch_loss": 0.015443 }, "checkpoint": "gemma4_e2b_v4input_topk_k32_L25_bnb4bit.pt" }, "input_split": { "n_neutral": 50, "n_v0": 50, "n_v3": 50, "neutral_vs_v0_block": { "label": "input_neutral_vs_v0", "d_max": 1.4726576805114746, "d_mean": 0.005705477204173803, "d_above_1": 11, "d_above_2": 0, "top10_ids": [ 2186, 5437, 3069, 2291, 6081, 2928, 3815, 2731, 4617, 1613 ], "top10_d": [ 1.4726576805114746, 1.43470299243927, 1.3176182508468628, 1.3077726364135742, 1.2976280450820923, 1.2607280015945435, 1.245164394378662, 1.2348037958145142, 1.151690125465393, 1.1124372482299805 ] }, "neutral_vs_v3_block": { "label": "input_neutral_vs_v3", "d_max": 2.9800453186035156, "d_mean": 0.007209507282823324, "d_above_1": 12, "d_above_2": 3, "top10_ids": [ 3815, 1703, 5437, 2731, 2291, 2186, 3686, 5465, 1931, 3069 ], "top10_d": [ 2.9800453186035156, 2.5605528354644775, 2.048558473587036, 1.9496744871139526, 1.8200923204421997, 1.6280057430267334, 1.5141710042953491, 1.4185123443603516, 1.3666783571243286, 1.3026963472366333 ] } }, "timestamp": "2026-04-21T06:53:37.744242+00:00" }