Add files using upload-large-folder tool

apply_query_grads_and_eval_reports.csv

@@ -0,0 +1,12 @@
+timestamp_utc,analysis_name,task,model_name,scores_type,bottom_k,margin,phase,learning_rate,reduce_fnc,task_loss,anthropic_mean_token_loss,initial_lr,final_lr,elapsed_seconds
+2025-08-09T09:09:22.769467+00:00,ethics_statichh-pythia-1.4b-sft-bf16,ethics,ncgc/statichh-pythia-1.4b-sft-bf16,pure_ekfac,100,0.1,bracket_search,0.0,mean,,1.798601,0.008,,894.405
+2025-08-09T09:24:38.693934+00:00,ethics_statichh-pythia-1.4b-sft-bf16,ethics,ncgc/statichh-pythia-1.4b-sft-bf16,pure_ekfac,100,0.1,bracket_search,0.008,mean,,1.801600,0.008,,898.468
+2025-08-09T09:39:47.962090+00:00,ethics_statichh-pythia-1.4b-sft-bf16,ethics,ncgc/statichh-pythia-1.4b-sft-bf16,pure_ekfac,100,0.1,bracket_search,0.04,mean,,1.826663,0.008,,899.171
+2025-08-09T09:54:56.134692+00:00,ethics_statichh-pythia-1.4b-sft-bf16,ethics,ncgc/statichh-pythia-1.4b-sft-bf16,pure_ekfac,100,0.1,binary_search,0.024,mean,,1.812035,0.008,,899.043
+2025-08-09T10:10:43.574276+00:00,ethics_statichh-pythia-1.4b-sft-bf16,ethics,ncgc/statichh-pythia-1.4b-sft-bf16,pure_ekfac,100,0.1,final_result,0.024,mean,-21.009991,1.812035,0.008,0.024,
+2025-08-09T10:41:20.180279+00:00,ethics_statichh-pythia-2.8b-sft-bf16,ethics,ncgc/statichh-pythia-2.8b-sft-bf16,pure_ekfac,100,0.1,bracket_search,0.0,mean,,1.718604,0.008,,1818.867
+2025-08-09T11:12:04.874055+00:00,ethics_statichh-pythia-2.8b-sft-bf16,ethics,ncgc/statichh-pythia-2.8b-sft-bf16,pure_ekfac,100,0.1,bracket_search,0.008,mean,,1.723822,0.008,,1819.604
+2025-08-09T11:42:36.940796+00:00,ethics_statichh-pythia-2.8b-sft-bf16,ethics,ncgc/statichh-pythia-2.8b-sft-bf16,pure_ekfac,100,0.1,bracket_search,0.04,mean,,1.799970,0.008,,1818.348
+2025-08-09T12:13:07.241727+00:00,ethics_statichh-pythia-2.8b-sft-bf16,ethics,ncgc/statichh-pythia-2.8b-sft-bf16,pure_ekfac,100,0.1,binary_search,0.024,mean,,1.750903,0.008,,1816.068
+2025-08-09T12:43:37.943954+00:00,ethics_statichh-pythia-2.8b-sft-bf16,ethics,ncgc/statichh-pythia-2.8b-sft-bf16,pure_ekfac,100,0.1,binary_search,0.016,mean,,1.734750,0.008,,1817.341
+2025-08-09T13:15:19.377791+00:00,ethics_statichh-pythia-2.8b-sft-bf16,ethics,ncgc/statichh-pythia-2.8b-sft-bf16,pure_ekfac,100,0.1,final_result,0.016,mean,-10.712186,1.734750,0.008,0.016,

bias_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/covariance_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 10000,
+    "indices": null
+}

bias_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/factor_arguments.json

@@ -0,0 +1,20 @@
+{
+    "strategy": "ekfac",
+    "use_empirical_fisher": false,
+    "amp_dtype": "torch.bfloat16",
+    "amp_scale": 65536.0,
+    "has_shared_parameters": false,
+    "covariance_max_examples": 100000,
+    "covariance_data_partitions": 1,
+    "covariance_module_partitions": 1,
+    "activation_covariance_dtype": "torch.bfloat16",
+    "gradient_covariance_dtype": "torch.bfloat16",
+    "eigendecomposition_dtype": "torch.float64",
+    "lambda_max_examples": 100000,
+    "lambda_data_partitions": 1,
+    "lambda_module_partitions": 1,
+    "use_iterative_lambda_aggregation": false,
+    "offload_activations_to_cpu": false,
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "lambda_dtype": "torch.bfloat16"
+}

bias_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/lambda_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 10000,
+    "indices": null
+}

bias_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/num_gradient_covariance_processed.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:262f6e940616054496a58e77f767a9fc85d0a54042507e12b324e32be39a4423
+size 1296

bias_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/factor_arguments.json

@@ -0,0 +1,20 @@
+{
+    "strategy": "ekfac",
+    "use_empirical_fisher": false,
+    "amp_dtype": "torch.bfloat16",
+    "amp_scale": 65536.0,
+    "has_shared_parameters": false,
+    "covariance_max_examples": 100000,
+    "covariance_data_partitions": 1,
+    "covariance_module_partitions": 1,
+    "activation_covariance_dtype": "torch.bfloat16",
+    "gradient_covariance_dtype": "torch.bfloat16",
+    "eigendecomposition_dtype": "torch.float64",
+    "lambda_max_examples": 100000,
+    "lambda_data_partitions": 1,
+    "lambda_module_partitions": 1,
+    "use_iterative_lambda_aggregation": false,
+    "offload_activations_to_cpu": false,
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "lambda_dtype": "torch.bfloat16"
+}

bias_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/pairwise_scores.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17797bd1eabf31eed7fdce9e884dfae46e738cdbfe58a54494228a09d763d585
+size 193192

bias_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/query_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 936,
+    "indices": null
+}

bias_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/score_arguments.json

@@ -0,0 +1,19 @@
+{
+    "damping_factor": 1e-08,
+    "amp_dtype": "torch.bfloat16",
+    "offload_activations_to_cpu": false,
+    "data_partitions": 1,
+    "module_partitions": 1,
+    "compute_per_module_scores": false,
+    "compute_per_token_scores": false,
+    "query_gradient_accumulation_steps": 1,
+    "query_gradient_low_rank": null,
+    "use_full_svd": false,
+    "aggregate_query_gradients": true,
+    "aggregate_train_gradients": false,
+    "use_measurement_for_self_influence": false,
+    "query_gradient_svd_dtype": "torch.float32",
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "precondition_dtype": "torch.bfloat16",
+    "score_dtype": "torch.bfloat16"
+}

bias_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/train_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 96256,
+    "indices": null
+}

bias_statichh-pythia-1.4b-sft-bf16/query_grads_pure_ekfac_bottom100/timing.json

@@ -0,0 +1,8 @@
+{
+  "model_name": "ncgc/statichh-pythia-1.4b-sft-bf16",
+  "analysis_name": "bias_statichh-pythia-1.4b-sft-bf16",
+  "scores_type": "pure_ekfac",
+  "bottom_k": 100,
+  "num_indices": 100,
+  "elapsed_seconds": 8.189496755599976
+}

bias_statichh-pythia-1.4b-sft-bf16/scores_ekfac_half/factor_arguments.json

@@ -0,0 +1,20 @@
+{
+    "strategy": "ekfac",
+    "use_empirical_fisher": false,
+    "amp_dtype": "torch.bfloat16",
+    "amp_scale": 65536.0,
+    "has_shared_parameters": false,
+    "covariance_max_examples": 100000,
+    "covariance_data_partitions": 1,
+    "covariance_module_partitions": 1,
+    "activation_covariance_dtype": "torch.bfloat16",
+    "gradient_covariance_dtype": "torch.bfloat16",
+    "eigendecomposition_dtype": "torch.float64",
+    "lambda_max_examples": 100000,
+    "lambda_data_partitions": 1,
+    "lambda_module_partitions": 1,
+    "use_iterative_lambda_aggregation": false,
+    "offload_activations_to_cpu": false,
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "lambda_dtype": "torch.bfloat16"
+}

bias_statichh-pythia-1.4b-sft-bf16/scores_ekfac_half/query_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 936,
+    "indices": null
+}

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/activation_covariance.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7a29713ab63105b9b184afaa2b96dac2b8d84367ed8168deabd8cb5653ac08f
+size 637650000

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/activation_eigenvalues.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8713974e0e928dd37581677a080e3dec83ede5e8dcb05a5acfb5e49e8e26a8f
+size 58400

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/activation_eigenvectors.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b1e70846951fbb5a1f2fd3834ff52da7f9cd32b0e635bded729a8b88a5422f1
+size 637650000

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/covariance_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 10000,
+    "indices": null
+}

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/factor_arguments.json

@@ -0,0 +1,20 @@
+{
+    "strategy": "ekfac",
+    "use_empirical_fisher": false,
+    "amp_dtype": "torch.bfloat16",
+    "amp_scale": 65536.0,
+    "has_shared_parameters": false,
+    "covariance_max_examples": 100000,
+    "covariance_data_partitions": 1,
+    "covariance_module_partitions": 1,
+    "activation_covariance_dtype": "torch.bfloat16",
+    "gradient_covariance_dtype": "torch.bfloat16",
+    "eigendecomposition_dtype": "torch.float64",
+    "lambda_max_examples": 100000,
+    "lambda_data_partitions": 1,
+    "lambda_module_partitions": 1,
+    "use_iterative_lambda_aggregation": false,
+    "offload_activations_to_cpu": false,
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "lambda_dtype": "torch.bfloat16"
+}

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/gradient_covariance.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b5f786acf49106a3fb8d6f5a5aa493d43c94e82588aeac4501c689ae7cbdaee
+size 905970768

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/gradient_eigenvalues.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebcbd5592e4a664984c8f3b95e9b452e28a768b0854791c40afc6098d7f73310
+size 74776

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/gradient_eigenvectors.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1259be57a6d759faff1ffb3fb6f485c80be764179cb4d893df2d3235e2da6503
+size 905970768

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/lambda_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 10000,
+    "indices": null
+}

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/lambda_matrix.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2e6ccc5df5b4b0ef41d0e956169e07dcd24919698761326a3dfb4157da2da31
+size 402728008

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/num_activation_covariance_processed.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8629f8669d8277cfb0d9091ffc44b14a4a21ccd39bbbfa63e0cc6cc94b37384a
+size 1040

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/num_gradient_covariance_processed.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92568f99e721482cb9f5661fdfa2562fb3d8ed47f883ce3cb25678bb1d9a8465
+size 1040

bias_statichh-pythia-6.9b-sft-bf16/factors_ekfac_half/num_lambda_processed.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d227cff54098ffcc8baf0f8127cde450470dfefc45e18b09e64d24a8b45b3e9
+size 1040

bias_statichh-pythia-6.9b-sft-bf16/query_grads_pure_ekfac_bottom100/aggregated_query_grads.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a14a4f25f548035d721f598f7b2e11522565e72548c35cdb4b6201fd6526c4a
+size 805454288

bias_statichh-pythia-6.9b-sft-bf16/query_grads_pure_ekfac_bottom100/timing.json

@@ -0,0 +1,8 @@
+{
+  "model_name": "ncgc/statichh-pythia-6.9b-sft-bf16",
+  "analysis_name": "bias_statichh-pythia-6.9b-sft-bf16",
+  "scores_type": "pure_ekfac",
+  "bottom_k": 100,
+  "num_indices": 100,
+  "elapsed_seconds": 27.0886549949646
+}

bias_statichh-pythia-6.9b-sft-bf16/scores_ekfac_half/factor_arguments.json

@@ -0,0 +1,20 @@
+{
+    "strategy": "ekfac",
+    "use_empirical_fisher": false,
+    "amp_dtype": "torch.bfloat16",
+    "amp_scale": 65536.0,
+    "has_shared_parameters": false,
+    "covariance_max_examples": 100000,
+    "covariance_data_partitions": 1,
+    "covariance_module_partitions": 1,
+    "activation_covariance_dtype": "torch.bfloat16",
+    "gradient_covariance_dtype": "torch.bfloat16",
+    "eigendecomposition_dtype": "torch.float64",
+    "lambda_max_examples": 100000,
+    "lambda_data_partitions": 1,
+    "lambda_module_partitions": 1,
+    "use_iterative_lambda_aggregation": false,
+    "offload_activations_to_cpu": false,
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "lambda_dtype": "torch.bfloat16"
+}

bias_statichh-pythia-6.9b-sft-bf16/scores_ekfac_half/pairwise_scores.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e83fb8c8fc762511e1b6e4fd9df0d98c25403988c02b0388211086c8146f7938
+size 193192

bias_statichh-pythia-6.9b-sft-bf16/scores_ekfac_half/query_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 936,
+    "indices": null
+}

bias_statichh-pythia-6.9b-sft-bf16/scores_ekfac_half/score_arguments.json

@@ -0,0 +1,19 @@
+{
+    "damping_factor": 1e-08,
+    "amp_dtype": "torch.bfloat16",
+    "offload_activations_to_cpu": false,
+    "data_partitions": 1,
+    "module_partitions": 1,
+    "compute_per_module_scores": false,
+    "compute_per_token_scores": false,
+    "query_gradient_accumulation_steps": 1,
+    "query_gradient_low_rank": null,
+    "use_full_svd": false,
+    "aggregate_query_gradients": true,
+    "aggregate_train_gradients": false,
+    "use_measurement_for_self_influence": false,
+    "query_gradient_svd_dtype": "torch.float32",
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "precondition_dtype": "torch.bfloat16",
+    "score_dtype": "torch.bfloat16"
+}

bias_statichh-pythia-6.9b-sft-bf16/scores_ekfac_half/train_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 96256,
+    "indices": null
+}

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/activation_covariance.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a6ffeca5cc1ad5f99fc7f8ea1a083a1207ee7068156b883f18f97b096382586
+size 188894582

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/activation_eigenvalues.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c5b21e2fd816ba39bdbc1627774f9c4d6ab65c372da1cb95e6fd1d1894db9ab
+size 37662

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/activation_eigenvectors.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2eda26e176a146a1f0e4d48dba827602ec541be20b445bd44986bc5ca08b7626
+size 188894582

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/covariance_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 10000,
+    "indices": null
+}

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/factor_arguments.json

@@ -0,0 +1,20 @@
+{
+    "strategy": "ekfac",
+    "use_empirical_fisher": false,
+    "amp_dtype": "torch.bfloat16",
+    "amp_scale": 65536.0,
+    "has_shared_parameters": false,
+    "covariance_max_examples": 100000,
+    "covariance_data_partitions": 1,
+    "covariance_module_partitions": 1,
+    "activation_covariance_dtype": "torch.bfloat16",
+    "gradient_covariance_dtype": "torch.bfloat16",
+    "eigendecomposition_dtype": "torch.float64",
+    "lambda_max_examples": 100000,
+    "lambda_data_partitions": 1,
+    "lambda_module_partitions": 1,
+    "use_iterative_lambda_aggregation": false,
+    "offload_activations_to_cpu": false,
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "lambda_dtype": "torch.bfloat16"
+}

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/gradient_covariance.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5facd479dc2f6cad903107a3f5b26e5d952e1bf9f9a47208ba07b1114ccbb49
+size 335545704

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/gradient_eigenvalues.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2c3c0f8e5023491d8d09e6a354a3ccb3d40d0b970874b0584a585b393b21059
+size 47384

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/gradient_eigenvectors.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:901c67bf1b756c5b9f991897a3081d18ad0f4226299d491dddb2bbbc03d241d9
+size 335545704

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/lambda_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 10000,
+    "indices": null
+}

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/lambda_matrix.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4317ef58af9b6c122880f3a153c58e7310ef7e0de03587b576c0d831c41c252f
+size 93590880

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/num_activation_covariance_processed.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d79037b28f31e4ed66f75668cc85a7e53a11f99deb533a29e70c15d13abf38a
+size 1296

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/num_gradient_covariance_processed.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1af79a1c7764da0b0e373066092f06e89370704a5b1b29fbd30b20cf4d37cfea
+size 1296

truth_statichh-Qwen-1.5B-sft-bf16/factors_ekfac_half/num_lambda_processed.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:101e45b366c3bd1c6e944d7b11b638e8d35cac2205a82286df26bc5e0c5a5357
+size 1296

truth_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/factor_arguments.json

@@ -0,0 +1,20 @@
+{
+    "strategy": "ekfac",
+    "use_empirical_fisher": false,
+    "amp_dtype": "torch.bfloat16",
+    "amp_scale": 65536.0,
+    "has_shared_parameters": false,
+    "covariance_max_examples": 100000,
+    "covariance_data_partitions": 1,
+    "covariance_module_partitions": 1,
+    "activation_covariance_dtype": "torch.bfloat16",
+    "gradient_covariance_dtype": "torch.bfloat16",
+    "eigendecomposition_dtype": "torch.float64",
+    "lambda_max_examples": 100000,
+    "lambda_data_partitions": 1,
+    "lambda_module_partitions": 1,
+    "use_iterative_lambda_aggregation": false,
+    "offload_activations_to_cpu": false,
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "lambda_dtype": "torch.bfloat16"
+}

truth_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/query_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 653,
+    "indices": null
+}

truth_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/score_arguments.json

@@ -0,0 +1,19 @@
+{
+    "damping_factor": 1e-08,
+    "amp_dtype": "torch.bfloat16",
+    "offload_activations_to_cpu": false,
+    "data_partitions": 1,
+    "module_partitions": 1,
+    "compute_per_module_scores": false,
+    "compute_per_token_scores": false,
+    "query_gradient_accumulation_steps": 1,
+    "query_gradient_low_rank": null,
+    "use_full_svd": false,
+    "aggregate_query_gradients": true,
+    "aggregate_train_gradients": false,
+    "use_measurement_for_self_influence": false,
+    "query_gradient_svd_dtype": "torch.float32",
+    "per_sample_gradient_dtype": "torch.bfloat16",
+    "precondition_dtype": "torch.bfloat16",
+    "score_dtype": "torch.bfloat16"
+}

truth_statichh-Qwen-1.5B-sft-bf16/scores_ekfac_half/train_dataset_metadata.json

@@ -0,0 +1,5 @@
+{
+    "type": "Dataset",
+    "dataset_size": 96256,
+    "indices": null
+}