| { |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "source": "OpenSAE/SAEBase v0.3.0 verified records", |
| "scope": "64 held-out controlled prompt groups per site", |
| "sites": [ |
| { |
| "site_id": "gemma270m:L5", |
| "model_id": "gemma270m", |
| "sae_id": "gemma270m_scope_resid", |
| "model_size": "270M", |
| "architecture": "jumprelu", |
| "layer": 5, |
| "hook_name": "blocks.5.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma270m_gemma270m_scope_resid_L5_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 29.59375, |
| "reconstruction_mse": 26.542215312258186, |
| "explained_variance": 0.9960876332784103, |
| "feature_ablation_logit_delta_l2": 544.4898433685303 |
| }, |
| { |
| "site_id": "gemma270m:L9", |
| "model_id": "gemma270m", |
| "sae_id": "gemma270m_scope_resid", |
| "model_size": "270M", |
| "architecture": "jumprelu", |
| "layer": 9, |
| "hook_name": "blocks.9.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma270m_gemma270m_scope_resid_L9_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 38.375, |
| "reconstruction_mse": 194.029170153415, |
| "explained_variance": 0.9956961126026779, |
| "feature_ablation_logit_delta_l2": 218.5887532234192 |
| }, |
| { |
| "site_id": "gemma270m:L12", |
| "model_id": "gemma270m", |
| "sae_id": "gemma270m_scope_resid", |
| "model_size": "270M", |
| "architecture": "jumprelu", |
| "layer": 12, |
| "hook_name": "blocks.12.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma270m_gemma270m_scope_resid_L12_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 73.828125, |
| "reconstruction_mse": 6211.399611460073, |
| "explained_variance": 0.9815030722617407, |
| "feature_ablation_logit_delta_l2": 192.075541973114 |
| }, |
| { |
| "site_id": "gemma1b:L7", |
| "model_id": "gemma1b", |
| "sae_id": "gemma1b_scope_resid", |
| "model_size": "1B", |
| "architecture": "jumprelu", |
| "layer": 7, |
| "hook_name": "blocks.7.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma1b_gemma1b_scope_resid_L7_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 44.203125, |
| "reconstruction_mse": 24.170324876026708, |
| "explained_variance": 0.9945018351843096, |
| "feature_ablation_logit_delta_l2": 998.0324802398682 |
| }, |
| { |
| "site_id": "gemma1b:L13", |
| "model_id": "gemma1b", |
| "sae_id": "gemma1b_scope_resid", |
| "model_size": "1B", |
| "architecture": "jumprelu", |
| "layer": 13, |
| "hook_name": "blocks.13.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma1b_gemma1b_scope_resid_L13_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 58.46875, |
| "reconstruction_mse": 392.15298906245454, |
| "explained_variance": 0.9946460049094926, |
| "feature_ablation_logit_delta_l2": 192.07367357611656 |
| }, |
| { |
| "site_id": "gemma1b:L17", |
| "model_id": "gemma1b", |
| "sae_id": "gemma1b_scope_resid", |
| "model_size": "1B", |
| "architecture": "jumprelu", |
| "layer": 17, |
| "hook_name": "blocks.17.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma1b_gemma1b_scope_resid_L17_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 57.421875, |
| "reconstruction_mse": 2120.874938720198, |
| "explained_variance": 0.9902167978901185, |
| "feature_ablation_logit_delta_l2": 281.59420013427734 |
| }, |
| { |
| "site_id": "gemma2_2b_base:L13", |
| "model_id": "gemma2_2b_base", |
| "sae_id": "gemma2_2b_pt_l25_canonical", |
| "model_size": "2B", |
| "architecture": "jumprelu", |
| "layer": 13, |
| "hook_name": "blocks.13.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma2_2b_base_gemma2_2b_pt_l25_canonical_L13_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 79.84375, |
| "reconstruction_mse": 1.8708208263731896, |
| "explained_variance": 0.8624282776229326, |
| "feature_ablation_logit_delta_l2": 1187.5063619613647 |
| }, |
| { |
| "site_id": "gemma2_2b_base:L25", |
| "model_id": "gemma2_2b_base", |
| "sae_id": "gemma2_2b_pt_l25_canonical", |
| "model_size": "2B", |
| "architecture": "jumprelu", |
| "layer": 25, |
| "hook_name": "blocks.25.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma2_2b_base_gemma2_2b_pt_l25_canonical_L25_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 88.234375, |
| "reconstruction_mse": 19.77263166884019, |
| "explained_variance": 0.8849783429040522, |
| "feature_ablation_logit_delta_l2": 691.7528381347656 |
| }, |
| { |
| "site_id": "qwen35_2b:L6", |
| "model_id": "qwen35_2b", |
| "sae_id": "qwen35_2b_resid_w32k", |
| "model_size": "2B", |
| "architecture": "topk", |
| "layer": 6, |
| "hook_name": "blocks.6.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_2b_qwen35_2b_resid_w32k_L6_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.006793501483693208, |
| "explained_variance": 0.46346274486470307, |
| "feature_ablation_logit_delta_l2": 56.52088910341263 |
| }, |
| { |
| "site_id": "qwen35_2b:L12", |
| "model_id": "qwen35_2b", |
| "sae_id": "qwen35_2b_resid_w32k", |
| "model_size": "2B", |
| "architecture": "topk", |
| "layer": 12, |
| "hook_name": "blocks.12.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_2b_qwen35_2b_resid_w32k_L12_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.009805408531902636, |
| "explained_variance": 0.5532504684400252, |
| "feature_ablation_logit_delta_l2": 69.88461792469025 |
| }, |
| { |
| "site_id": "qwen35_2b:L15", |
| "model_id": "qwen35_2b", |
| "sae_id": "qwen35_2b_resid_w32k", |
| "model_size": "2B", |
| "architecture": "topk", |
| "layer": 15, |
| "hook_name": "blocks.15.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_2b_qwen35_2b_resid_w32k_L15_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.022683608212980867, |
| "explained_variance": 0.6552050741283569, |
| "feature_ablation_logit_delta_l2": 434.9815306663513 |
| }, |
| { |
| "site_id": "qwen35_2b:L20", |
| "model_id": "qwen35_2b", |
| "sae_id": "qwen35_2b_resid_w32k", |
| "model_size": "2B", |
| "architecture": "topk", |
| "layer": 20, |
| "hook_name": "blocks.20.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_2b_qwen35_2b_resid_w32k_L20_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.12875008172159408, |
| "explained_variance": 0.6393610397161591, |
| "feature_ablation_logit_delta_l2": 88.30731463432312 |
| }, |
| { |
| "site_id": "qwen3_8b:L9", |
| "model_id": "qwen3_8b", |
| "sae_id": "qwen3_8b_resid_w64k", |
| "model_size": "8B", |
| "architecture": "topk", |
| "layer": 9, |
| "hook_name": "blocks.9.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_8b_qwen3_8b_resid_w64k_L9_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.4320439064708014, |
| "explained_variance": 0.31170907396943487, |
| "feature_ablation_logit_delta_l2": 90.98629534244537 |
| }, |
| { |
| "site_id": "qwen3_8b:L18", |
| "model_id": "qwen3_8b", |
| "sae_id": "qwen3_8b_resid_w64k", |
| "model_size": "8B", |
| "architecture": "topk", |
| "layer": 18, |
| "hook_name": "blocks.18.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_8b_qwen3_8b_resid_w64k_L18_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 1.2545259425764421, |
| "explained_variance": 0.4155404086706985, |
| "feature_ablation_logit_delta_l2": 80.89625352621078 |
| }, |
| { |
| "site_id": "qwen3_8b:L23", |
| "model_id": "qwen3_8b", |
| "sae_id": "qwen3_8b_resid_w64k", |
| "model_size": "8B", |
| "architecture": "topk", |
| "layer": 23, |
| "hook_name": "blocks.23.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_8b_qwen3_8b_resid_w64k_L23_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 5.515160429610223, |
| "explained_variance": 0.40572943601426786, |
| "feature_ablation_logit_delta_l2": 183.35777175426483 |
| }, |
| { |
| "site_id": "qwen3_8b:L31", |
| "model_id": "qwen3_8b", |
| "sae_id": "qwen3_8b_resid_w64k", |
| "model_size": "8B", |
| "architecture": "topk", |
| "layer": 31, |
| "hook_name": "blocks.31.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_8b_qwen3_8b_resid_w64k_L31_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 104.54906565321956, |
| "explained_variance": 0.3955223965348901, |
| "feature_ablation_logit_delta_l2": 427.6726031303406 |
| }, |
| { |
| "site_id": "llama31_8b:L8", |
| "model_id": "llama31_8b", |
| "sae_id": "llama31_8b_lxr_8x", |
| "model_size": "8B", |
| "architecture": "jumprelu", |
| "layer": 8, |
| "hook_name": "blocks.8.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/llama31_8b_llama31_8b_lxr_8x_L8_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 28.046875, |
| "reconstruction_mse": 0.004055736768313844, |
| "explained_variance": 0.5513006794989936, |
| "feature_ablation_logit_delta_l2": 21.845480531454086 |
| }, |
| { |
| "site_id": "llama31_8b:L16", |
| "model_id": "llama31_8b", |
| "sae_id": "llama31_8b_lxr_8x", |
| "model_size": "8B", |
| "architecture": "jumprelu", |
| "layer": 16, |
| "hook_name": "blocks.16.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/llama31_8b_llama31_8b_lxr_8x_L16_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 30.859375, |
| "reconstruction_mse": 0.012340238014570772, |
| "explained_variance": 0.6725650422908716, |
| "feature_ablation_logit_delta_l2": 54.130552887916565 |
| }, |
| { |
| "site_id": "llama31_8b:L21", |
| "model_id": "llama31_8b", |
| "sae_id": "llama31_8b_lxr_8x", |
| "model_size": "8B", |
| "architecture": "jumprelu", |
| "layer": 21, |
| "hook_name": "blocks.21.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/llama31_8b_llama31_8b_lxr_8x_L21_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 33.71875, |
| "reconstruction_mse": 0.03148705278742039, |
| "explained_variance": 0.7491312626018514, |
| "feature_ablation_logit_delta_l2": 127.93421411514282 |
| }, |
| { |
| "site_id": "llama31_8b:L27", |
| "model_id": "llama31_8b", |
| "sae_id": "llama31_8b_lxr_8x", |
| "model_size": "8B", |
| "architecture": "jumprelu", |
| "layer": 27, |
| "hook_name": "blocks.27.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/llama31_8b_llama31_8b_lxr_8x_L27_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 25.5625, |
| "reconstruction_mse": 0.0830698530201267, |
| "explained_variance": 0.7886849246530725, |
| "feature_ablation_logit_delta_l2": 453.17595958709717 |
| }, |
| { |
| "site_id": "pythia410m:L4", |
| "model_id": "pythia410m", |
| "sae_id": "pythia410m_mlp_topk", |
| "model_size": "410M", |
| "architecture": "topk", |
| "layer": 4, |
| "hook_name": "blocks.4.hook_mlp_out", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/pythia410m_pythia410m_mlp_topk_L4_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 32, |
| "reconstruction_mse": 0.025108318325979897, |
| "explained_variance": 0.19371843102737688, |
| "feature_ablation_logit_delta_l2": 1.4841452287510037 |
| }, |
| { |
| "site_id": "pythia410m:L12", |
| "model_id": "pythia410m", |
| "sae_id": "pythia410m_mlp_topk", |
| "model_size": "410M", |
| "architecture": "topk", |
| "layer": 12, |
| "hook_name": "blocks.12.hook_mlp_out", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/pythia410m_pythia410m_mlp_topk_L12_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 32, |
| "reconstruction_mse": 0.04636237055879008, |
| "explained_variance": -0.06422395528236455, |
| "feature_ablation_logit_delta_l2": 0.731195330619812 |
| }, |
| { |
| "site_id": "pythia410m:L20", |
| "model_id": "pythia410m", |
| "sae_id": "pythia410m_mlp_topk", |
| "model_size": "410M", |
| "architecture": "topk", |
| "layer": 20, |
| "hook_name": "blocks.20.hook_mlp_out", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/pythia410m_pythia410m_mlp_topk_L20_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 32, |
| "reconstruction_mse": 0.2914138840317801, |
| "explained_variance": 0.2665335035612107, |
| "feature_ablation_logit_delta_l2": 18.107948787510395 |
| }, |
| { |
| "site_id": "qwen3_1_7b_base:L7", |
| "model_id": "qwen3_1_7b_base", |
| "sae_id": "qwen3_1_7b_base_resid_w32k", |
| "model_size": "1.7B", |
| "architecture": "topk", |
| "layer": 7, |
| "hook_name": "blocks.7.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_1_7b_base_qwen3_1_7b_base_resid_w32k_L7_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.9558896948083689, |
| "explained_variance": 0.29946944044652013, |
| "feature_ablation_logit_delta_l2": 58.22856664657593 |
| }, |
| { |
| "site_id": "qwen3_1_7b_base:L14", |
| "model_id": "qwen3_1_7b_base", |
| "sae_id": "qwen3_1_7b_base_resid_w32k", |
| "model_size": "1.7B", |
| "architecture": "topk", |
| "layer": 14, |
| "hook_name": "blocks.14.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_1_7b_base_qwen3_1_7b_base_resid_w32k_L14_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 8.740651689114003, |
| "explained_variance": 0.3422633758354428, |
| "feature_ablation_logit_delta_l2": 103.53831255435944 |
| }, |
| { |
| "site_id": "qwen3_1_7b_base:L18", |
| "model_id": "qwen3_1_7b_base", |
| "sae_id": "qwen3_1_7b_base_resid_w32k", |
| "model_size": "1.7B", |
| "architecture": "topk", |
| "layer": 18, |
| "hook_name": "blocks.18.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_1_7b_base_qwen3_1_7b_base_resid_w32k_L18_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 55.69111087201516, |
| "explained_variance": 0.5916587266615272, |
| "feature_ablation_logit_delta_l2": 142.1756579875946 |
| }, |
| { |
| "site_id": "qwen3_1_7b_base:L23", |
| "model_id": "qwen3_1_7b_base", |
| "sae_id": "qwen3_1_7b_base_resid_w32k", |
| "model_size": "1.7B", |
| "architecture": "topk", |
| "layer": 23, |
| "hook_name": "blocks.23.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_1_7b_base_qwen3_1_7b_base_resid_w32k_L23_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 920.5619560351674, |
| "explained_variance": 0.5140207115832733, |
| "feature_ablation_logit_delta_l2": 230.97397637367249 |
| }, |
| { |
| "site_id": "r1_distill_qwen15b:L20", |
| "model_id": "r1_distill_qwen15b", |
| "sae_id": "r1_distill_qwen15b_resid_l20", |
| "model_size": "1.5B", |
| "architecture": "topk", |
| "layer": 20, |
| "hook_name": "blocks.20.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/r1_distill_qwen15b_r1_distill_qwen15b_resid_l20_L20_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 32, |
| "reconstruction_mse": 3.873100533740864, |
| "explained_variance": 0.7829561668350811, |
| "feature_ablation_logit_delta_l2": 379.62211990356445 |
| }, |
| { |
| "site_id": "gemma3_4b_pt:L9", |
| "model_id": "gemma3_4b_pt", |
| "sae_id": "gemma3_4b_pt_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 9, |
| "hook_name": "blocks.9.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_pt_gemma3_4b_pt_scope_resid_L9_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 54.390625, |
| "reconstruction_mse": 55.37114540220211, |
| "explained_variance": 0.9980620877887129, |
| "feature_ablation_logit_delta_l2": 82.46295690536499 |
| }, |
| { |
| "site_id": "gemma3_4b_pt:L17", |
| "model_id": "gemma3_4b_pt", |
| "sae_id": "gemma3_4b_pt_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 17, |
| "hook_name": "blocks.17.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_pt_gemma3_4b_pt_scope_resid_L17_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 51.453125, |
| "reconstruction_mse": 745.7371491667323, |
| "explained_variance": 0.9980640516682647, |
| "feature_ablation_logit_delta_l2": 39.621697932481766 |
| }, |
| { |
| "site_id": "gemma3_4b_pt:L22", |
| "model_id": "gemma3_4b_pt", |
| "sae_id": "gemma3_4b_pt_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 22, |
| "hook_name": "blocks.22.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_pt_gemma3_4b_pt_scope_resid_L22_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 55.328125, |
| "reconstruction_mse": 3340.604917947747, |
| "explained_variance": 0.995293750208509, |
| "feature_ablation_logit_delta_l2": 185.91756331920624 |
| }, |
| { |
| "site_id": "gemma3_4b_pt:L29", |
| "model_id": "gemma3_4b_pt", |
| "sae_id": "gemma3_4b_pt_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 29, |
| "hook_name": "blocks.29.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_pt_gemma3_4b_pt_scope_resid_L29_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 67.609375, |
| "reconstruction_mse": 16670.329024322135, |
| "explained_variance": 0.9924005801734385, |
| "feature_ablation_logit_delta_l2": 134.11140948534012 |
| }, |
| { |
| "site_id": "gemma3_4b_it:L9", |
| "model_id": "gemma3_4b_it", |
| "sae_id": "gemma3_4b_it_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 9, |
| "hook_name": "blocks.9.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_it_gemma3_4b_it_scope_resid_L9_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 49.46875, |
| "reconstruction_mse": 59.02352188537665, |
| "explained_variance": 0.9975901200976052, |
| "feature_ablation_logit_delta_l2": 252.38312005996704 |
| }, |
| { |
| "site_id": "gemma3_4b_it:L17", |
| "model_id": "gemma3_4b_it", |
| "sae_id": "gemma3_4b_it_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 17, |
| "hook_name": "blocks.17.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_it_gemma3_4b_it_scope_resid_L17_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 51.390625, |
| "reconstruction_mse": 1053.2008350506942, |
| "explained_variance": 0.9974764196568326, |
| "feature_ablation_logit_delta_l2": 0 |
| }, |
| { |
| "site_id": "gemma3_4b_it:L22", |
| "model_id": "gemma3_4b_it", |
| "sae_id": "gemma3_4b_it_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 22, |
| "hook_name": "blocks.22.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_it_gemma3_4b_it_scope_resid_L22_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 59.546875, |
| "reconstruction_mse": 3539.259447411133, |
| "explained_variance": 0.9904257167550933, |
| "feature_ablation_logit_delta_l2": 304.3531255722046 |
| }, |
| { |
| "site_id": "gemma3_4b_it:L29", |
| "model_id": "gemma3_4b_it", |
| "sae_id": "gemma3_4b_it_scope_resid", |
| "model_size": "4B", |
| "architecture": "jumprelu", |
| "layer": 29, |
| "hook_name": "blocks.29.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma3_4b_it_gemma3_4b_it_scope_resid_L29_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 56.578125, |
| "reconstruction_mse": 13861.33423753159, |
| "explained_variance": 0.9838387190907543, |
| "feature_ablation_logit_delta_l2": 87.25914072990417 |
| }, |
| { |
| "site_id": "mistral7b_instruct:L16", |
| "model_id": "mistral7b_instruct", |
| "sae_id": "mistral7b_instruct_tc_l16_131k", |
| "model_size": "7B", |
| "architecture": "topk", |
| "layer": 16, |
| "hook_name": "blocks.16.hook_mlp_out", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/mistral7b_instruct_mistral7b_instruct_tc_l16_131k_L16_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 128, |
| "reconstruction_mse": 0.0017195014873718525, |
| "explained_variance": 0.229585189758012, |
| "feature_ablation_logit_delta_l2": 4.043052937835455 |
| }, |
| { |
| "site_id": "gemma2_9b_pt:L9", |
| "model_id": "gemma2_9b_pt", |
| "sae_id": "gemma2_9b_pt_scope_resid", |
| "model_size": "9B", |
| "architecture": "jumprelu", |
| "layer": 9, |
| "hook_name": "blocks.9.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma2_9b_pt_gemma2_9b_pt_scope_resid_L9_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 40.984375, |
| "reconstruction_mse": 0.5731250118357542, |
| "explained_variance": 0.7633360054820594, |
| "feature_ablation_logit_delta_l2": 1630.5595483779907 |
| }, |
| { |
| "site_id": "gemma2_9b_pt:L20", |
| "model_id": "gemma2_9b_pt", |
| "sae_id": "gemma2_9b_pt_scope_resid", |
| "model_size": "9B", |
| "architecture": "jumprelu", |
| "layer": 20, |
| "hook_name": "blocks.20.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma2_9b_pt_gemma2_9b_pt_scope_resid_L20_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 53.203125, |
| "reconstruction_mse": 2.470728176167096, |
| "explained_variance": 0.8284526015476957, |
| "feature_ablation_logit_delta_l2": 1109.2622900009155 |
| }, |
| { |
| "site_id": "gemma2_9b_pt:L31", |
| "model_id": "gemma2_9b_pt", |
| "sae_id": "gemma2_9b_pt_scope_resid", |
| "model_size": "9B", |
| "architecture": "jumprelu", |
| "layer": 31, |
| "hook_name": "blocks.31.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma2_9b_pt_gemma2_9b_pt_scope_resid_L31_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 37.9375, |
| "reconstruction_mse": 7.0037165051193995, |
| "explained_variance": 0.8832148493687838, |
| "feature_ablation_logit_delta_l2": 882.8945960998535 |
| }, |
| { |
| "site_id": "qwen35_9b:L8", |
| "model_id": "qwen35_9b", |
| "sae_id": "qwen35_9b_resid_w64k", |
| "model_size": "9B", |
| "architecture": "topk", |
| "layer": 8, |
| "hook_name": "blocks.8.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_9b_qwen35_9b_resid_w64k_L8_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.09302263526948515, |
| "explained_variance": 0.2821747926370852, |
| "feature_ablation_logit_delta_l2": 97.08006256818771 |
| }, |
| { |
| "site_id": "qwen35_9b:L16", |
| "model_id": "qwen35_9b", |
| "sae_id": "qwen35_9b_resid_w64k", |
| "model_size": "9B", |
| "architecture": "topk", |
| "layer": 16, |
| "hook_name": "blocks.16.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_9b_qwen35_9b_resid_w64k_L16_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.25773795690173984, |
| "explained_variance": 0.339508538673662, |
| "feature_ablation_logit_delta_l2": 200.36437845230103 |
| }, |
| { |
| "site_id": "qwen35_9b:L21", |
| "model_id": "qwen35_9b", |
| "sae_id": "qwen35_9b_resid_w64k", |
| "model_size": "9B", |
| "architecture": "topk", |
| "layer": 21, |
| "hook_name": "blocks.21.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_9b_qwen35_9b_resid_w64k_L21_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 0.5343900374019106, |
| "explained_variance": 0.6449151143444183, |
| "feature_ablation_logit_delta_l2": 197.24622082710266 |
| }, |
| { |
| "site_id": "qwen35_9b:L27", |
| "model_id": "qwen35_9b", |
| "sae_id": "qwen35_9b_resid_w64k", |
| "model_size": "9B", |
| "architecture": "topk", |
| "layer": 27, |
| "hook_name": "blocks.27.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_9b_qwen35_9b_resid_w64k_L27_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 50, |
| "reconstruction_mse": 2.2175642009104575, |
| "explained_variance": 0.5774464680435034, |
| "feature_ablation_logit_delta_l2": 225.19099187850952 |
| }, |
| { |
| "site_id": "r1_distill_llama8b:L23", |
| "model_id": "r1_distill_llama8b", |
| "sae_id": "r1_distill_llama8b_llamascope_l23", |
| "model_size": "8B", |
| "architecture": "jumprelu", |
| "layer": 23, |
| "hook_name": "blocks.23.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/r1_distill_llama8b_r1_distill_llama8b_llamascope_l23_L23_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 8.125, |
| "reconstruction_mse": 0.08067539281105633, |
| "explained_variance": 0.45463792560988686, |
| "feature_ablation_logit_delta_l2": 81.63063311576843 |
| }, |
| { |
| "site_id": "qwen3_1_7b:L14", |
| "model_id": "qwen3_1_7b", |
| "sae_id": "qwen3_1_7b_ak_resid_l14_65k", |
| "model_size": "1.7B", |
| "architecture": "batchtopk", |
| "layer": 14, |
| "hook_name": "blocks.14.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen3_1_7b_qwen3_1_7b_ak_resid_l14_65k_L14_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 69.984375, |
| "reconstruction_mse": 2.0263488726996317, |
| "explained_variance": 0.8517183866628193, |
| "feature_ablation_logit_delta_l2": 512.9422640800476 |
| }, |
| { |
| "site_id": "gemma2_2b_it:L12", |
| "model_id": "gemma2_2b_it", |
| "sae_id": "gemma2_2b_it_matryoshka_l12_32k", |
| "model_size": "2B", |
| "architecture": "matryoshka_batchtopk", |
| "layer": 12, |
| "hook_name": "blocks.12.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/gemma2_2b_it_gemma2_2b_it_matryoshka_l12_32k_L12_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 153.3125, |
| "reconstruction_mse": 3.3529621877163858, |
| "explained_variance": 0.7824819296171213, |
| "feature_ablation_logit_delta_l2": 262.2079429626465 |
| }, |
| { |
| "site_id": "qwen35_4b:L15", |
| "model_id": "qwen35_4b", |
| "sae_id": "qwen35_4b_matryoshka_l15_65k", |
| "model_size": "4B", |
| "architecture": "matryoshka_batchtopk", |
| "layer": 15, |
| "hook_name": "blocks.15.hook_resid_post", |
| "compatibility": "passed", |
| "benchmark": "passed", |
| "reliability": "passed", |
| "evaluation_groups": 64, |
| "protocol": "controlled_heldout_groups", |
| "report_url": "https://github.com/OpenSAE/SAEBase/blob/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279/conf/coverage/stage_d_evidence/reports/qwen35_4b_qwen35_4b_matryoshka_l15_65k_L15_m3m4.json", |
| "source_commit": "d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279", |
| "mean_l0": 100.03125, |
| "reconstruction_mse": 0.007906571600660514, |
| "explained_variance": 0.8096178164397694, |
| "feature_ablation_logit_delta_l2": 633.4192380905151 |
| } |
| ] |
| } |
|
|
