Add λ-sweep iso-L0 family: 5 λ + 3-seed at λ=1.35 (Pythia-160M W_U d8192)

pythia-160m/W_U/lambda-sweep/d8192/lam0p40_seed0.config.json

@@ -0,0 +1,119 @@
+{
+  "public_label": "Pythia-160M W_U crosscoder, \u03bb-sweep iso-L0 (\u03bb=0.4), 32 snapshots, d_sae=8192, seed 0",
+  "config": {
+    "sae_type": "crosscoder",
+    "d_model": 768,
+    "expansion_factor": 10.6667,
+    "use_decoder_bias": true,
+    "act_fn": "jumprelu",
+    "norm_activation": "inference",
+    "sparsity_include_decoder_norm": true,
+    "top_k": 50,
+    "use_triton_kernel": false,
+    "sparsity_threshold_for_triton_spmm_kernel": 0.996,
+    "jumprelu_threshold_window": 4.0,
+    "hook_points": [
+      "wu_step_0",
+      "wu_step_1",
+      "wu_step_2",
+      "wu_step_3",
+      "wu_step_4",
+      "wu_step_5",
+      "wu_step_6",
+      "wu_step_7",
+      "wu_step_8",
+      "wu_step_9",
+      "wu_step_10",
+      "wu_step_11",
+      "wu_step_12",
+      "wu_step_13",
+      "wu_step_14",
+      "wu_step_15",
+      "wu_step_16",
+      "wu_step_17",
+      "wu_step_18",
+      "wu_step_19",
+      "wu_step_20",
+      "wu_step_21",
+      "wu_step_22",
+      "wu_step_23",
+      "wu_step_24",
+      "wu_step_25",
+      "wu_step_26",
+      "wu_step_27",
+      "wu_step_28",
+      "wu_step_29",
+      "wu_step_30",
+      "wu_step_31"
+    ]
+  },
+  "training": {
+    "lr": 5e-05,
+    "l1_coefficient": 0.4,
+    "frequency_scale": 0.01,
+    "tanh_stretch_coefficient": 1.0,
+    "jumprelu_lr_factor": 0.1,
+    "init_threshold": 0.1,
+    "decoder_transpose_init": 1.0,
+    "amp_dtype": "fp32",
+    "optimizer": "adam",
+    "n_epochs": 300,
+    "batch_size": 1024,
+    "expansion_factor": 10.6667,
+    "d_sae": 8192,
+    "input_preprocess": "center_scale",
+    "l1_warmup_fraction": 0.1,
+    "lr_warmup_fraction": 0.0,
+    "lr_decay_fraction": 0.0,
+    "auxk_coefficient": 0.03125,
+    "auxk_k": 512,
+    "dead_window_steps": 100
+  },
+  "quality": {
+    "explained_variance": 0.7479858952216084,
+    "reconstruction_mse": 0.25193381249904634,
+    "mean_l0": 160.0768243408203,
+    "dead_rate": 0.0,
+    "d_sae": 8192,
+    "n_snapshots": 32,
+    "n_rows": 50304
+  },
+  "steps": [
+    0,
+    1,
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128,
+    256,
+    512,
+    1000,
+    2000,
+    3000,
+    4000,
+    5000,
+    6000,
+    7000,
+    8000,
+    9000,
+    14000,
+    21000,
+    27000,
+    34000,
+    47000,
+    61000,
+    75000,
+    89000,
+    102000,
+    116000,
+    130000,
+    143000
+  ],
+  "model_name": "EleutherAI/pythia-160m",
+  "seed": 0,
+  "kind": "cross-snapshot",
+  "preprocess_mode": "center_scale"
+}

pythia-160m/W_U/lambda-sweep/d8192/lam0p40_seed0.md

@@ -0,0 +1,39 @@
+# Pythia-160M W_U crosscoder, λ-sweep iso-L0 (λ=0.4), 32 snapshots, d_sae=8192, seed 0
+
+Companion artifact to *Parameter-trajectory crosscoders for vocabulary readout
+evolution* (NeurIPS 2026). Member of the **λ-sweep iso-L0 family**: a 7-point
+sparsity sweep at d_sae=8192, sharing the released training recipe except for
+the L1 coefficient (and adding `auxk_coefficient=0.03125` to prevent dead-feature
+collapse at high λ). Three replicate seeds at λ=1.35 give a variance estimate.
+
+## Quality
+- Explained variance: 0.7479858952216084
+- Reconstruction MSE: 0.25193381249904634
+- Mean L0: 160.0768243408203
+- Dead-feature rate: 0.0
+- d_sae: 8192
+- n_snapshots: 32
+
+## Training recipe (key hparams)
+- Activation: jumprelu
+- L1 coefficient: 0.4
+- AuxK coefficient: 0.03125
+- AuxK k: 512
+- Frequency penalty scale: 0.01
+- Tanh-stretch coefficient: 1.0
+- LR: 5e-05
+- JumpReLU LR multiplier: 0.1
+- Init threshold: 0.1
+- Optimizer: adam
+- Epochs: 300
+- Batch size: 1024
+- Preprocess: center_scale
+- Seed: 0
+
+## Loading
+```python
+from safetensors.torch import load_file
+import json
+weights = load_file("lam0p40_seed0.safetensors")
+config = json.load(open("lam0p40_seed0.config.json"))
+```

pythia-160m/W_U/lambda-sweep/d8192/lam0p40_seed0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29c5a8e6cd17cb39a9c383371362636fc649c5bac693add96e692e621e4f9c4a
+size 1611792824

pythia-160m/W_U/lambda-sweep/d8192/lam1p00_seed0.config.json

@@ -0,0 +1,119 @@
+{
+  "public_label": "Pythia-160M W_U crosscoder, \u03bb-sweep iso-L0 (\u03bb=1.0), 32 snapshots, d_sae=8192, seed 0",
+  "config": {
+    "sae_type": "crosscoder",
+    "d_model": 768,
+    "expansion_factor": 10.6667,
+    "use_decoder_bias": true,
+    "act_fn": "jumprelu",
+    "norm_activation": "inference",
+    "sparsity_include_decoder_norm": true,
+    "top_k": 50,
+    "use_triton_kernel": false,
+    "sparsity_threshold_for_triton_spmm_kernel": 0.996,
+    "jumprelu_threshold_window": 4.0,
+    "hook_points": [
+      "wu_step_0",
+      "wu_step_1",
+      "wu_step_2",
+      "wu_step_3",
+      "wu_step_4",
+      "wu_step_5",
+      "wu_step_6",
+      "wu_step_7",
+      "wu_step_8",
+      "wu_step_9",
+      "wu_step_10",
+      "wu_step_11",
+      "wu_step_12",
+      "wu_step_13",
+      "wu_step_14",
+      "wu_step_15",
+      "wu_step_16",
+      "wu_step_17",
+      "wu_step_18",
+      "wu_step_19",
+      "wu_step_20",
+      "wu_step_21",
+      "wu_step_22",
+      "wu_step_23",
+      "wu_step_24",
+      "wu_step_25",
+      "wu_step_26",
+      "wu_step_27",
+      "wu_step_28",
+      "wu_step_29",
+      "wu_step_30",
+      "wu_step_31"
+    ]
+  },
+  "training": {
+    "lr": 5e-05,
+    "l1_coefficient": 1.0,
+    "frequency_scale": 0.01,
+    "tanh_stretch_coefficient": 1.0,
+    "jumprelu_lr_factor": 0.1,
+    "init_threshold": 0.1,
+    "decoder_transpose_init": 1.0,
+    "amp_dtype": "fp32",
+    "optimizer": "adam",
+    "n_epochs": 300,
+    "batch_size": 1024,
+    "expansion_factor": 10.6667,
+    "d_sae": 8192,
+    "input_preprocess": "center_scale",
+    "l1_warmup_fraction": 0.1,
+    "lr_warmup_fraction": 0.0,
+    "lr_decay_fraction": 0.0,
+    "auxk_coefficient": 0.03125,
+    "auxk_k": 512,
+    "dead_window_steps": 100
+  },
+  "quality": {
+    "explained_variance": 0.6323049070911215,
+    "reconstruction_mse": 0.36757794439792635,
+    "mean_l0": 58.49865539550781,
+    "dead_rate": 0.0,
+    "d_sae": 8192,
+    "n_snapshots": 32,
+    "n_rows": 50304
+  },
+  "steps": [
+    0,
+    1,
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128,
+    256,
+    512,
+    1000,
+    2000,
+    3000,
+    4000,
+    5000,
+    6000,
+    7000,
+    8000,
+    9000,
+    14000,
+    21000,
+    27000,
+    34000,
+    47000,
+    61000,
+    75000,
+    89000,
+    102000,
+    116000,
+    130000,
+    143000
+  ],
+  "model_name": "EleutherAI/pythia-160m",
+  "seed": 0,
+  "kind": "cross-snapshot",
+  "preprocess_mode": "center_scale"
+}

pythia-160m/W_U/lambda-sweep/d8192/lam1p00_seed0.md

@@ -0,0 +1,39 @@
+# Pythia-160M W_U crosscoder, λ-sweep iso-L0 (λ=1.0), 32 snapshots, d_sae=8192, seed 0
+
+Companion artifact to *Parameter-trajectory crosscoders for vocabulary readout
+evolution* (NeurIPS 2026). Member of the **λ-sweep iso-L0 family**: a 7-point
+sparsity sweep at d_sae=8192, sharing the released training recipe except for
+the L1 coefficient (and adding `auxk_coefficient=0.03125` to prevent dead-feature
+collapse at high λ). Three replicate seeds at λ=1.35 give a variance estimate.
+
+## Quality
+- Explained variance: 0.6323049070911215
+- Reconstruction MSE: 0.36757794439792635
+- Mean L0: 58.49865539550781
+- Dead-feature rate: 0.0
+- d_sae: 8192
+- n_snapshots: 32
+
+## Training recipe (key hparams)
+- Activation: jumprelu
+- L1 coefficient: 1.0
+- AuxK coefficient: 0.03125
+- AuxK k: 512
+- Frequency penalty scale: 0.01
+- Tanh-stretch coefficient: 1.0
+- LR: 5e-05
+- JumpReLU LR multiplier: 0.1
+- Init threshold: 0.1
+- Optimizer: adam
+- Epochs: 300
+- Batch size: 1024
+- Preprocess: center_scale
+- Seed: 0
+
+## Loading
+```python
+from safetensors.torch import load_file
+import json
+weights = load_file("lam1p00_seed0.safetensors")
+config = json.load(open("lam1p00_seed0.config.json"))
+```

pythia-160m/W_U/lambda-sweep/d8192/lam1p00_seed0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e60e0f2f23729140050503cfb60496d791808d0764b2907a0a7a3eb6559bd182
+size 1611792824

pythia-160m/W_U/lambda-sweep/d8192/lam1p20_seed0.config.json

@@ -0,0 +1,119 @@
+{
+  "public_label": "Pythia-160M W_U crosscoder, \u03bb-sweep iso-L0 (\u03bb=1.2), 32 snapshots, d_sae=8192, seed 0",
+  "config": {
+    "sae_type": "crosscoder",
+    "d_model": 768,
+    "expansion_factor": 10.6667,
+    "use_decoder_bias": true,
+    "act_fn": "jumprelu",
+    "norm_activation": "inference",
+    "sparsity_include_decoder_norm": true,
+    "top_k": 50,
+    "use_triton_kernel": false,
+    "sparsity_threshold_for_triton_spmm_kernel": 0.996,
+    "jumprelu_threshold_window": 4.0,
+    "hook_points": [
+      "wu_step_0",
+      "wu_step_1",
+      "wu_step_2",
+      "wu_step_3",
+      "wu_step_4",
+      "wu_step_5",
+      "wu_step_6",
+      "wu_step_7",
+      "wu_step_8",
+      "wu_step_9",
+      "wu_step_10",
+      "wu_step_11",
+      "wu_step_12",
+      "wu_step_13",
+      "wu_step_14",
+      "wu_step_15",
+      "wu_step_16",
+      "wu_step_17",
+      "wu_step_18",
+      "wu_step_19",
+      "wu_step_20",
+      "wu_step_21",
+      "wu_step_22",
+      "wu_step_23",
+      "wu_step_24",
+      "wu_step_25",
+      "wu_step_26",
+      "wu_step_27",
+      "wu_step_28",
+      "wu_step_29",
+      "wu_step_30",
+      "wu_step_31"
+    ]
+  },
+  "training": {
+    "lr": 5e-05,
+    "l1_coefficient": 1.2,
+    "frequency_scale": 0.01,
+    "tanh_stretch_coefficient": 1.0,
+    "jumprelu_lr_factor": 0.1,
+    "init_threshold": 0.1,
+    "decoder_transpose_init": 1.0,
+    "amp_dtype": "fp32",
+    "optimizer": "adam",
+    "n_epochs": 300,
+    "batch_size": 1024,
+    "expansion_factor": 10.6667,
+    "d_sae": 8192,
+    "input_preprocess": "center_scale",
+    "l1_warmup_fraction": 0.1,
+    "lr_warmup_fraction": 0.0,
+    "lr_decay_fraction": 0.0,
+    "auxk_coefficient": 0.03125,
+    "auxk_k": 512,
+    "dead_window_steps": 100
+  },
+  "quality": {
+    "explained_variance": 0.6027639222075235,
+    "reconstruction_mse": 0.39710951745510104,
+    "mean_l0": 45.41338317871094,
+    "dead_rate": 0.0,
+    "d_sae": 8192,
+    "n_snapshots": 32,
+    "n_rows": 50304
+  },
+  "steps": [
+    0,
+    1,
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128,
+    256,
+    512,
+    1000,
+    2000,
+    3000,
+    4000,
+    5000,
+    6000,
+    7000,
+    8000,
+    9000,
+    14000,
+    21000,
+    27000,
+    34000,
+    47000,
+    61000,
+    75000,
+    89000,
+    102000,
+    116000,
+    130000,
+    143000
+  ],
+  "model_name": "EleutherAI/pythia-160m",
+  "seed": 0,
+  "kind": "cross-snapshot",
+  "preprocess_mode": "center_scale"
+}

pythia-160m/W_U/lambda-sweep/d8192/lam1p20_seed0.md

@@ -0,0 +1,39 @@
+# Pythia-160M W_U crosscoder, λ-sweep iso-L0 (λ=1.2), 32 snapshots, d_sae=8192, seed 0
+
+Companion artifact to *Parameter-trajectory crosscoders for vocabulary readout
+evolution* (NeurIPS 2026). Member of the **λ-sweep iso-L0 family**: a 7-point
+sparsity sweep at d_sae=8192, sharing the released training recipe except for
+the L1 coefficient (and adding `auxk_coefficient=0.03125` to prevent dead-feature
+collapse at high λ). Three replicate seeds at λ=1.35 give a variance estimate.
+
+## Quality
+- Explained variance: 0.6027639222075235
+- Reconstruction MSE: 0.39710951745510104
+- Mean L0: 45.41338317871094
+- Dead-feature rate: 0.0
+- d_sae: 8192
+- n_snapshots: 32
+
+## Training recipe (key hparams)
+- Activation: jumprelu
+- L1 coefficient: 1.2
+- AuxK coefficient: 0.03125
+- AuxK k: 512
+- Frequency penalty scale: 0.01
+- Tanh-stretch coefficient: 1.0
+- LR: 5e-05
+- JumpReLU LR multiplier: 0.1
+- Init threshold: 0.1
+- Optimizer: adam
+- Epochs: 300
+- Batch size: 1024
+- Preprocess: center_scale
+- Seed: 0
+
+## Loading
+```python
+from safetensors.torch import load_file
+import json
+weights = load_file("lam1p20_seed0.safetensors")
+config = json.load(open("lam1p20_seed0.config.json"))
+```

pythia-160m/W_U/lambda-sweep/d8192/lam1p20_seed0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e03cc5543a16e184f7da8993aabaa67baad9e57ed19f36440b956f3791b2f45f
+size 1611792824

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed0.config.json

@@ -0,0 +1,119 @@
+{
+  "public_label": "Pythia-160M W_U crosscoder, \u03bb-sweep iso-L0 (\u03bb=1.35), 32 snapshots, d_sae=8192, seed 0",
+  "config": {
+    "sae_type": "crosscoder",
+    "d_model": 768,
+    "expansion_factor": 10.6667,
+    "use_decoder_bias": true,
+    "act_fn": "jumprelu",
+    "norm_activation": "inference",
+    "sparsity_include_decoder_norm": true,
+    "top_k": 50,
+    "use_triton_kernel": false,
+    "sparsity_threshold_for_triton_spmm_kernel": 0.996,
+    "jumprelu_threshold_window": 4.0,
+    "hook_points": [
+      "wu_step_0",
+      "wu_step_1",
+      "wu_step_2",
+      "wu_step_3",
+      "wu_step_4",
+      "wu_step_5",
+      "wu_step_6",
+      "wu_step_7",
+      "wu_step_8",
+      "wu_step_9",
+      "wu_step_10",
+      "wu_step_11",
+      "wu_step_12",
+      "wu_step_13",
+      "wu_step_14",
+      "wu_step_15",
+      "wu_step_16",
+      "wu_step_17",
+      "wu_step_18",
+      "wu_step_19",
+      "wu_step_20",
+      "wu_step_21",
+      "wu_step_22",
+      "wu_step_23",
+      "wu_step_24",
+      "wu_step_25",
+      "wu_step_26",
+      "wu_step_27",
+      "wu_step_28",
+      "wu_step_29",
+      "wu_step_30",
+      "wu_step_31"
+    ]
+  },
+  "training": {
+    "lr": 5e-05,
+    "l1_coefficient": 1.35,
+    "frequency_scale": 0.01,
+    "tanh_stretch_coefficient": 1.0,
+    "jumprelu_lr_factor": 0.1,
+    "init_threshold": 0.1,
+    "decoder_transpose_init": 1.0,
+    "amp_dtype": "fp32",
+    "optimizer": "adam",
+    "n_epochs": 300,
+    "batch_size": 1024,
+    "expansion_factor": 10.6667,
+    "d_sae": 8192,
+    "input_preprocess": "center_scale",
+    "l1_warmup_fraction": 0.1,
+    "lr_warmup_fraction": 0.0,
+    "lr_decay_fraction": 0.0,
+    "auxk_coefficient": 0.03125,
+    "auxk_k": 512,
+    "dead_window_steps": 100
+  },
+  "quality": {
+    "explained_variance": 0.5819729868726617,
+    "reconstruction_mse": 0.4178938287496567,
+    "mean_l0": 37.971342163085936,
+    "dead_rate": 0.0,
+    "d_sae": 8192,
+    "n_snapshots": 32,
+    "n_rows": 50304
+  },
+  "steps": [
+    0,
+    1,
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128,
+    256,
+    512,
+    1000,
+    2000,
+    3000,
+    4000,
+    5000,
+    6000,
+    7000,
+    8000,
+    9000,
+    14000,
+    21000,
+    27000,
+    34000,
+    47000,
+    61000,
+    75000,
+    89000,
+    102000,
+    116000,
+    130000,
+    143000
+  ],
+  "model_name": "EleutherAI/pythia-160m",
+  "seed": 0,
+  "kind": "cross-snapshot",
+  "preprocess_mode": "center_scale"
+}

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed0.md

@@ -0,0 +1,39 @@
+# Pythia-160M W_U crosscoder, λ-sweep iso-L0 (λ=1.35), 32 snapshots, d_sae=8192, seed 0
+
+Companion artifact to *Parameter-trajectory crosscoders for vocabulary readout
+evolution* (NeurIPS 2026). Member of the **λ-sweep iso-L0 family**: a 7-point
+sparsity sweep at d_sae=8192, sharing the released training recipe except for
+the L1 coefficient (and adding `auxk_coefficient=0.03125` to prevent dead-feature
+collapse at high λ). Three replicate seeds at λ=1.35 give a variance estimate.
+
+## Quality
+- Explained variance: 0.5819729868726617
+- Reconstruction MSE: 0.4178938287496567
+- Mean L0: 37.971342163085936
+- Dead-feature rate: 0.0
+- d_sae: 8192
+- n_snapshots: 32
+
+## Training recipe (key hparams)
+- Activation: jumprelu
+- L1 coefficient: 1.35
+- AuxK coefficient: 0.03125
+- AuxK k: 512
+- Frequency penalty scale: 0.01
+- Tanh-stretch coefficient: 1.0
+- LR: 5e-05
+- JumpReLU LR multiplier: 0.1
+- Init threshold: 0.1
+- Optimizer: adam
+- Epochs: 300
+- Batch size: 1024
+- Preprocess: center_scale
+- Seed: 0
+
+## Loading
+```python
+from safetensors.torch import load_file
+import json
+weights = load_file("lam1p35_seed0.safetensors")
+config = json.load(open("lam1p35_seed0.config.json"))
+```

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca937c8bc0270d32dbc4af90aba71b50131f4101efdfb546293c61e918f4d73f
+size 1611792824

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed1.config.json

@@ -0,0 +1,119 @@
+{
+  "public_label": "Pythia-160M W_U crosscoder, \u03bb-sweep iso-L0 (\u03bb=1.35), 32 snapshots, d_sae=8192, seed 1",
+  "config": {
+    "sae_type": "crosscoder",
+    "d_model": 768,
+    "expansion_factor": 10.6667,
+    "use_decoder_bias": true,
+    "act_fn": "jumprelu",
+    "norm_activation": "inference",
+    "sparsity_include_decoder_norm": true,
+    "top_k": 50,
+    "use_triton_kernel": false,
+    "sparsity_threshold_for_triton_spmm_kernel": 0.996,
+    "jumprelu_threshold_window": 4.0,
+    "hook_points": [
+      "wu_step_0",
+      "wu_step_1",
+      "wu_step_2",
+      "wu_step_3",
+      "wu_step_4",
+      "wu_step_5",
+      "wu_step_6",
+      "wu_step_7",
+      "wu_step_8",
+      "wu_step_9",
+      "wu_step_10",
+      "wu_step_11",
+      "wu_step_12",
+      "wu_step_13",
+      "wu_step_14",
+      "wu_step_15",
+      "wu_step_16",
+      "wu_step_17",
+      "wu_step_18",
+      "wu_step_19",
+      "wu_step_20",
+      "wu_step_21",
+      "wu_step_22",
+      "wu_step_23",
+      "wu_step_24",
+      "wu_step_25",
+      "wu_step_26",
+      "wu_step_27",
+      "wu_step_28",
+      "wu_step_29",
+      "wu_step_30",
+      "wu_step_31"
+    ]
+  },
+  "training": {
+    "lr": 5e-05,
+    "l1_coefficient": 1.35,
+    "frequency_scale": 0.01,
+    "tanh_stretch_coefficient": 1.0,
+    "jumprelu_lr_factor": 0.1,
+    "init_threshold": 0.1,
+    "decoder_transpose_init": 1.0,
+    "amp_dtype": "fp32",
+    "optimizer": "adam",
+    "n_epochs": 300,
+    "batch_size": 1024,
+    "expansion_factor": 10.6667,
+    "d_sae": 8192,
+    "input_preprocess": "center_scale",
+    "l1_warmup_fraction": 0.1,
+    "lr_warmup_fraction": 0.0,
+    "lr_decay_fraction": 0.0,
+    "auxk_coefficient": 0.03125,
+    "auxk_k": 512,
+    "dead_window_steps": 100
+  },
+  "quality": {
+    "explained_variance": 0.581652004193779,
+    "reconstruction_mse": 0.4182147091627121,
+    "mean_l0": 37.79685302734375,
+    "dead_rate": 0.0,
+    "d_sae": 8192,
+    "n_snapshots": 32,
+    "n_rows": 50304
+  },
+  "steps": [
+    0,
+    1,
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128,
+    256,
+    512,
+    1000,
+    2000,
+    3000,
+    4000,
+    5000,
+    6000,
+    7000,
+    8000,
+    9000,
+    14000,
+    21000,
+    27000,
+    34000,
+    47000,
+    61000,
+    75000,
+    89000,
+    102000,
+    116000,
+    130000,
+    143000
+  ],
+  "model_name": "EleutherAI/pythia-160m",
+  "seed": 1,
+  "kind": "cross-snapshot",
+  "preprocess_mode": "center_scale"
+}

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed1.md

@@ -0,0 +1,39 @@
+# Pythia-160M W_U crosscoder, λ-sweep iso-L0 (λ=1.35), 32 snapshots, d_sae=8192, seed 1
+
+Companion artifact to *Parameter-trajectory crosscoders for vocabulary readout
+evolution* (NeurIPS 2026). Member of the **λ-sweep iso-L0 family**: a 7-point
+sparsity sweep at d_sae=8192, sharing the released training recipe except for
+the L1 coefficient (and adding `auxk_coefficient=0.03125` to prevent dead-feature
+collapse at high λ). Three replicate seeds at λ=1.35 give a variance estimate.
+
+## Quality
+- Explained variance: 0.581652004193779
+- Reconstruction MSE: 0.4182147091627121
+- Mean L0: 37.79685302734375
+- Dead-feature rate: 0.0
+- d_sae: 8192
+- n_snapshots: 32
+
+## Training recipe (key hparams)
+- Activation: jumprelu
+- L1 coefficient: 1.35
+- AuxK coefficient: 0.03125
+- AuxK k: 512
+- Frequency penalty scale: 0.01
+- Tanh-stretch coefficient: 1.0
+- LR: 5e-05
+- JumpReLU LR multiplier: 0.1
+- Init threshold: 0.1
+- Optimizer: adam
+- Epochs: 300
+- Batch size: 1024
+- Preprocess: center_scale
+- Seed: 1
+
+## Loading
+```python
+from safetensors.torch import load_file
+import json
+weights = load_file("lam1p35_seed1.safetensors")
+config = json.load(open("lam1p35_seed1.config.json"))
+```

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:185d4f40c64f8a5956c9aec82ca7cd8b612b1b6d20468067463a08dd172c23fb
+size 1611792824

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed2.config.json

@@ -0,0 +1,119 @@
+{
+  "public_label": "Pythia-160M W_U crosscoder, \u03bb-sweep iso-L0 (\u03bb=1.35), 32 snapshots, d_sae=8192, seed 2",
+  "config": {
+    "sae_type": "crosscoder",
+    "d_model": 768,
+    "expansion_factor": 10.6667,
+    "use_decoder_bias": true,
+    "act_fn": "jumprelu",
+    "norm_activation": "inference",
+    "sparsity_include_decoder_norm": true,
+    "top_k": 50,
+    "use_triton_kernel": false,
+    "sparsity_threshold_for_triton_spmm_kernel": 0.996,
+    "jumprelu_threshold_window": 4.0,
+    "hook_points": [
+      "wu_step_0",
+      "wu_step_1",
+      "wu_step_2",
+      "wu_step_3",
+      "wu_step_4",
+      "wu_step_5",
+      "wu_step_6",
+      "wu_step_7",
+      "wu_step_8",
+      "wu_step_9",
+      "wu_step_10",
+      "wu_step_11",
+      "wu_step_12",
+      "wu_step_13",
+      "wu_step_14",
+      "wu_step_15",
+      "wu_step_16",
+      "wu_step_17",
+      "wu_step_18",
+      "wu_step_19",
+      "wu_step_20",
+      "wu_step_21",
+      "wu_step_22",
+      "wu_step_23",
+      "wu_step_24",
+      "wu_step_25",
+      "wu_step_26",
+      "wu_step_27",
+      "wu_step_28",
+      "wu_step_29",
+      "wu_step_30",
+      "wu_step_31"
+    ]
+  },
+  "training": {
+    "lr": 5e-05,
+    "l1_coefficient": 1.35,
+    "frequency_scale": 0.01,
+    "tanh_stretch_coefficient": 1.0,
+    "jumprelu_lr_factor": 0.1,
+    "init_threshold": 0.1,
+    "decoder_transpose_init": 1.0,
+    "amp_dtype": "fp32",
+    "optimizer": "adam",
+    "n_epochs": 300,
+    "batch_size": 1024,
+    "expansion_factor": 10.6667,
+    "d_sae": 8192,
+    "input_preprocess": "center_scale",
+    "l1_warmup_fraction": 0.1,
+    "lr_warmup_fraction": 0.0,
+    "lr_decay_fraction": 0.0,
+    "auxk_coefficient": 0.03125,
+    "auxk_k": 512,
+    "dead_window_steps": 100
+  },
+  "quality": {
+    "explained_variance": 0.582366028895355,
+    "reconstruction_mse": 0.4175009119510651,
+    "mean_l0": 37.980972900390626,
+    "dead_rate": 0.0,
+    "d_sae": 8192,
+    "n_snapshots": 32,
+    "n_rows": 50304
+  },
+  "steps": [
+    0,
+    1,
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128,
+    256,
+    512,
+    1000,
+    2000,
+    3000,
+    4000,
+    5000,
+    6000,
+    7000,
+    8000,
+    9000,
+    14000,
+    21000,
+    27000,
+    34000,
+    47000,
+    61000,
+    75000,
+    89000,
+    102000,
+    116000,
+    130000,
+    143000
+  ],
+  "model_name": "EleutherAI/pythia-160m",
+  "seed": 2,
+  "kind": "cross-snapshot",
+  "preprocess_mode": "center_scale"
+}

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed2.md

@@ -0,0 +1,39 @@
+# Pythia-160M W_U crosscoder, λ-sweep iso-L0 (λ=1.35), 32 snapshots, d_sae=8192, seed 2
+
+Companion artifact to *Parameter-trajectory crosscoders for vocabulary readout
+evolution* (NeurIPS 2026). Member of the **λ-sweep iso-L0 family**: a 7-point
+sparsity sweep at d_sae=8192, sharing the released training recipe except for
+the L1 coefficient (and adding `auxk_coefficient=0.03125` to prevent dead-feature
+collapse at high λ). Three replicate seeds at λ=1.35 give a variance estimate.
+
+## Quality
+- Explained variance: 0.582366028895355
+- Reconstruction MSE: 0.4175009119510651
+- Mean L0: 37.980972900390626
+- Dead-feature rate: 0.0
+- d_sae: 8192
+- n_snapshots: 32
+
+## Training recipe (key hparams)
+- Activation: jumprelu
+- L1 coefficient: 1.35
+- AuxK coefficient: 0.03125
+- AuxK k: 512
+- Frequency penalty scale: 0.01
+- Tanh-stretch coefficient: 1.0
+- LR: 5e-05
+- JumpReLU LR multiplier: 0.1
+- Init threshold: 0.1
+- Optimizer: adam
+- Epochs: 300
+- Batch size: 1024
+- Preprocess: center_scale
+- Seed: 2
+
+## Loading
+```python
+from safetensors.torch import load_file
+import json
+weights = load_file("lam1p35_seed2.safetensors")
+config = json.load(open("lam1p35_seed2.config.json"))
+```

pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed2.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:727983bad97a97e20ff7089f967fc4b0071a45f1d9bf181bae2381f2f63d4e14
+size 1611792824

pythia-160m/W_U/lambda-sweep/d8192/lam1p80_seed0.config.json

@@ -0,0 +1,119 @@
+{
+  "public_label": "Pythia-160M W_U crosscoder, \u03bb-sweep iso-L0 (\u03bb=1.8), 32 snapshots, d_sae=8192, seed 0",
+  "config": {
+    "sae_type": "crosscoder",
+    "d_model": 768,
+    "expansion_factor": 10.6667,
+    "use_decoder_bias": true,
+    "act_fn": "jumprelu",
+    "norm_activation": "inference",
+    "sparsity_include_decoder_norm": true,
+    "top_k": 50,
+    "use_triton_kernel": false,
+    "sparsity_threshold_for_triton_spmm_kernel": 0.996,
+    "jumprelu_threshold_window": 4.0,
+    "hook_points": [
+      "wu_step_0",
+      "wu_step_1",
+      "wu_step_2",
+      "wu_step_3",
+      "wu_step_4",
+      "wu_step_5",
+      "wu_step_6",
+      "wu_step_7",
+      "wu_step_8",
+      "wu_step_9",
+      "wu_step_10",
+      "wu_step_11",
+      "wu_step_12",
+      "wu_step_13",
+      "wu_step_14",
+      "wu_step_15",
+      "wu_step_16",
+      "wu_step_17",
+      "wu_step_18",
+      "wu_step_19",
+      "wu_step_20",
+      "wu_step_21",
+      "wu_step_22",
+      "wu_step_23",
+      "wu_step_24",
+      "wu_step_25",
+      "wu_step_26",
+      "wu_step_27",
+      "wu_step_28",
+      "wu_step_29",
+      "wu_step_30",
+      "wu_step_31"
+    ]
+  },
+  "training": {
+    "lr": 5e-05,
+    "l1_coefficient": 1.8,
+    "frequency_scale": 0.01,
+    "tanh_stretch_coefficient": 1.0,
+    "jumprelu_lr_factor": 0.1,
+    "init_threshold": 0.1,
+    "decoder_transpose_init": 1.0,
+    "amp_dtype": "fp32",
+    "optimizer": "adam",
+    "n_epochs": 300,
+    "batch_size": 1024,
+    "expansion_factor": 10.6667,
+    "d_sae": 8192,
+    "input_preprocess": "center_scale",
+    "l1_warmup_fraction": 0.1,
+    "lr_warmup_fraction": 0.0,
+    "lr_decay_fraction": 0.0,
+    "auxk_coefficient": 0.03125,
+    "auxk_k": 512,
+    "dead_window_steps": 100
+  },
+  "quality": {
+    "explained_variance": 0.527703732469858,
+    "reconstruction_mse": 0.4721457928419113,
+    "mean_l0": 23.033436279296875,
+    "dead_rate": 0.0,
+    "d_sae": 8192,
+    "n_snapshots": 32,
+    "n_rows": 50304
+  },
+  "steps": [
+    0,
+    1,
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128,
+    256,
+    512,
+    1000,
+    2000,
+    3000,
+    4000,
+    5000,
+    6000,
+    7000,
+    8000,
+    9000,
+    14000,
+    21000,
+    27000,
+    34000,
+    47000,
+    61000,
+    75000,
+    89000,
+    102000,
+    116000,
+    130000,
+    143000
+  ],
+  "model_name": "EleutherAI/pythia-160m",
+  "seed": 0,
+  "kind": "cross-snapshot",
+  "preprocess_mode": "center_scale"
+}

pythia-160m/W_U/lambda-sweep/d8192/lam1p80_seed0.md

@@ -0,0 +1,39 @@
+# Pythia-160M W_U crosscoder, λ-sweep iso-L0 (λ=1.8), 32 snapshots, d_sae=8192, seed 0
+
+Companion artifact to *Parameter-trajectory crosscoders for vocabulary readout
+evolution* (NeurIPS 2026). Member of the **λ-sweep iso-L0 family**: a 7-point
+sparsity sweep at d_sae=8192, sharing the released training recipe except for
+the L1 coefficient (and adding `auxk_coefficient=0.03125` to prevent dead-feature
+collapse at high λ). Three replicate seeds at λ=1.35 give a variance estimate.
+
+## Quality
+- Explained variance: 0.527703732469858
+- Reconstruction MSE: 0.4721457928419113
+- Mean L0: 23.033436279296875
+- Dead-feature rate: 0.0
+- d_sae: 8192
+- n_snapshots: 32
+
+## Training recipe (key hparams)
+- Activation: jumprelu
+- L1 coefficient: 1.8
+- AuxK coefficient: 0.03125
+- AuxK k: 512
+- Frequency penalty scale: 0.01
+- Tanh-stretch coefficient: 1.0
+- LR: 5e-05
+- JumpReLU LR multiplier: 0.1
+- Init threshold: 0.1
+- Optimizer: adam
+- Epochs: 300
+- Batch size: 1024
+- Preprocess: center_scale
+- Seed: 0
+
+## Loading
+```python
+from safetensors.torch import load_file
+import json
+weights = load_file("lam1p80_seed0.safetensors")
+config = json.load(open("lam1p80_seed0.config.json"))
+```

pythia-160m/W_U/lambda-sweep/d8192/lam1p80_seed0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0dba1e20536dad7105f47bf0afe97c6b908caba3bf8ba4a2ddd1fd49a4471e2c
+size 1611792824