Datasets:

AverageMetaheuristicsEnjoyer
/

mobe-factor-cache-gpt-oss-20b-b8

Tags:

mobe

moe-compression

License:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Preview

Duplicate

Split (1)

train

The full dataset viewer is not available (click to read why). Only showing a preview of the rows.

The dataset generation failed because of a cast error

Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 3 new columns ({'error_type', 'status', 'error_message'}) and 296 missing columns ({'mlp_in_affected_blocks_compression_ratio', 'norm_ratio_after_mlp_method_over_og_num_tokens', 'norm_delta_layer_method_median', 'rel_diff_norm_after_mha_og_vs_method_num_tokens', 'superweight_rel_error_mean', 'norm_delta_mha_og_mean', 'diff_norm_after_mlp_og_vs_method_mean', 'norm_after_layer_method_mean', 'diff_norm_after_layer_og_vs_method_num_tokens', 'cos_delta_mlp_to_layer_og_vs_method_std', 'cos_after_mha_residual_og_vs_method_mean', 'angle_deg_after_mha_og_vs_method_mean', 'moe_experts_in_affected_blocks_dense_bits', 'diff_norm_after_layer_og_vs_method_median', 'ppl_wikitext2', 'rel_diff_norm_after_mlp_og_vs_method_std', 'benchmarked_with_dense_reconstruction', 'rel_diff_norm_delta_layer_og_vs_method_std', 'norm_ratio_delta_mlp_to_layer_method_over_og_mean', 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_std', 'diff_norm_after_mha_residual_og_vs_method_std', 'target_layers', 'recipe', 'rel_diff_norm_before_layer_og_vs_method_std', 'overall_with_embeddings_dense_bits', 'norm_before_layer_og_median', 'cos_after_mha_residual_og_vs_method_std', 'diff_norm_delta_mha_og_vs_method_std', 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_median', 'mha_in_affected_blocks_dense_bits', 'model_arch_key', 'affected_blocks_compression_ratio', 'overall_no_embeddings_compression_ratio', 'angle_deg_after_mlp_og_vs_method_std', 'norm_ratio_after_layer_method_over_og_num_tokens', 'diff_norm_delta_mha_og_vs_method_num_tokens', 'norm_after_layer_og_mean', 'angle_deg_delta_mha_og_vs_method_std
...
orm_after_mha_og_vs_method_median', 'norm_delta_mlp_to_layer_og_mean', 'norm_after_mlp_method_num_tokens', 'cos_before_layer_og_vs_method_std', 'cos_delta_layer_og_vs_method_median', 'target_layer_bits', 'cos_after_mha_og_vs_method_num_tokens', 'mlp_in_affected_blocks_dense_bits', 'diff_norm_after_mha_og_vs_method_mean', 'norm_ratio_delta_mha_method_over_og_mean', 'norm_delta_mlp_to_layer_method_median', 'norm_after_mha_method_mean', 'norm_before_layer_method_std', 'rel_diff_norm_after_mha_residual_og_vs_method_num_tokens', 'norm_after_mha_og_mean', 'cos_before_layer_og_vs_method_num_tokens', 'cos_before_layer_og_vs_method_median', 'norm_delta_mlp_to_layer_method_mean', 'diff_norm_after_mlp_og_vs_method_std', 'overall_no_embeddings_dense_bits', 'diff_norm_after_mha_og_vs_method_num_tokens', 'lm_eval_attempted', 'norm_ratio_delta_mha_method_over_og_median', 'torch_compile_skip_reason', 'cos_after_mha_residual_og_vs_method_num_tokens', 'rel_diff_norm_after_layer_og_vs_method_num_tokens', 'rel_diff_norm_after_layer_og_vs_method_mean', 'norm_delta_mha_method_std', 'resume_key', 'superweight_abs_error_mean', 'cos_after_layer_og_vs_method_num_tokens', 'angle_deg_delta_mha_og_vs_method_median', 'norm_ratio_before_layer_method_over_og_mean', 'norm_ratio_delta_mlp_to_layer_method_over_og_num_tokens', 'diff_norm_after_mlp_og_vs_method_num_tokens', 'norm_after_mha_residual_og_std', 'affected_modules_storage_bits', 'cos_after_mlp_og_vs_method_std', 'diff_norm_after_mha_og_vs_method_std'}).

This happened while the csv dataset builder was generating data using

hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8/openai_gpt-oss-20b__partial_results__status.csv (at revision 5ea6bfa03283e398864b7526ef72b8ffa083895b), ['hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8@5ea6bfa03283e398864b7526ef72b8ffa083895b/gpt_oss_20b_mobe_b8_results.csv', 'hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8@5ea6bfa03283e398864b7526ef72b8ffa083895b/openai_gpt-oss-20b__partial_results.csv', 'hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8@5ea6bfa03283e398864b7526ef72b8ffa083895b/openai_gpt-oss-20b__partial_results__status.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1837, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
                  pa_table = table_cast(pa_table, self._schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              model_name: string
              target_set: string
              method_label: string
              status: string
              error_type: double
              error_message: double
              rank: double
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 1111
              to
              {'target_set': Value('string'), 'model_arch_key': Value('string'), 'model_family': Value('string'), 'model_name': Value('string'), 'method_label': Value('string'), 'rank': Value('float64'), 'target_modules': Value('string'), 'target_layers': Value('string'), 'storage_records': Value('string'), 'affected_modules_dense_bits': Value('int64'), 'affected_modules_storage_bits': Value('int64'), 'affected_modules_compression_ratio': Value('float64'), 'mha_in_affected_blocks_dense_bits': Value('float64'), 'mha_in_affected_blocks_storage_bits': Value('float64'), 'mha_in_affected_blocks_compression_ratio': Value('float64'), 'mlp_in_affected_blocks_dense_bits': Value('float64'), 'mlp_in_affected_blocks_storage_bits': Value('float64'), 'mlp_in_affected_blocks_compression_ratio': Value('float64'), 'affected_blocks_dense_bits': Value('int64'), 'affected_blocks_storage_bits': Value('int64'), 'affected_blocks_compression_ratio': Value('float64'), 'overall_no_embeddings_dense_bits': Value('int64'), 'overall_no_embeddings_storage_bits': Value('int64'), 'overall_no_embeddings_compression_ratio': Value('float64'), 'overall_with_embeddings_dense_bits': Value('int64'), 'overall_with_embeddings_storage_bits': Value('int64'), 'overall_with_embeddings_compression_ratio': Value('float64'), 'target_layer_bits': Value('int64'), 'overall_model_bits': Value('int64'), 'target_layer_compression_ratio': Value('float64'), 'overall_model_compression_ratio': Value('float64'), 'superweight_abs_error_mean': Value(
              ...
              _og_vs_method_median': Value('float64'), 'angle_deg_delta_mlp_to_layer_og_vs_method_std': Value('float64'), 'angle_deg_delta_mlp_to_layer_og_vs_method_num_tokens': Value('float64'), 'norm_delta_mlp_to_layer_og_mean': Value('float64'), 'norm_delta_mlp_to_layer_og_median': Value('float64'), 'norm_delta_mlp_to_layer_og_std': Value('float64'), 'norm_delta_mlp_to_layer_og_num_tokens': Value('float64'), 'norm_delta_mlp_to_layer_method_mean': Value('float64'), 'norm_delta_mlp_to_layer_method_median': Value('float64'), 'norm_delta_mlp_to_layer_method_std': Value('float64'), 'norm_delta_mlp_to_layer_method_num_tokens': Value('float64'), 'norm_ratio_delta_mlp_to_layer_method_over_og_mean': Value('float64'), 'norm_ratio_delta_mlp_to_layer_method_over_og_median': Value('float64'), 'norm_ratio_delta_mlp_to_layer_method_over_og_std': Value('float64'), 'norm_ratio_delta_mlp_to_layer_method_over_og_num_tokens': Value('float64'), 'diff_norm_delta_mlp_to_layer_og_vs_method_mean': Value('float64'), 'diff_norm_delta_mlp_to_layer_og_vs_method_median': Value('float64'), 'diff_norm_delta_mlp_to_layer_og_vs_method_std': Value('float64'), 'diff_norm_delta_mlp_to_layer_og_vs_method_num_tokens': Value('float64'), 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_mean': Value('float64'), 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_median': Value('float64'), 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_std': Value('float64'), 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_num_tokens': Value('float64')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1839, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 3 new columns ({'error_type', 'status', 'error_message'}) and 296 missing columns ({'mlp_in_affected_blocks_compression_ratio', 'norm_ratio_after_mlp_method_over_og_num_tokens', 'norm_delta_layer_method_median', 'rel_diff_norm_after_mha_og_vs_method_num_tokens', 'superweight_rel_error_mean', 'norm_delta_mha_og_mean', 'diff_norm_after_mlp_og_vs_method_mean', 'norm_after_layer_method_mean', 'diff_norm_after_layer_og_vs_method_num_tokens', 'cos_delta_mlp_to_layer_og_vs_method_std', 'cos_after_mha_residual_og_vs_method_mean', 'angle_deg_after_mha_og_vs_method_mean', 'moe_experts_in_affected_blocks_dense_bits', 'diff_norm_after_layer_og_vs_method_median', 'ppl_wikitext2', 'rel_diff_norm_after_mlp_og_vs_method_std', 'benchmarked_with_dense_reconstruction', 'rel_diff_norm_delta_layer_og_vs_method_std', 'norm_ratio_delta_mlp_to_layer_method_over_og_mean', 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_std', 'diff_norm_after_mha_residual_og_vs_method_std', 'target_layers', 'recipe', 'rel_diff_norm_before_layer_og_vs_method_std', 'overall_with_embeddings_dense_bits', 'norm_before_layer_og_median', 'cos_after_mha_residual_og_vs_method_std', 'diff_norm_delta_mha_og_vs_method_std', 'rel_diff_norm_delta_mlp_to_layer_og_vs_method_median', 'mha_in_affected_blocks_dense_bits', 'model_arch_key', 'affected_blocks_compression_ratio', 'overall_no_embeddings_compression_ratio', 'angle_deg_after_mlp_og_vs_method_std', 'norm_ratio_after_layer_method_over_og_num_tokens', 'diff_norm_delta_mha_og_vs_method_num_tokens', 'norm_after_layer_og_mean', 'angle_deg_delta_mha_og_vs_method_std
              ...
              orm_after_mha_og_vs_method_median', 'norm_delta_mlp_to_layer_og_mean', 'norm_after_mlp_method_num_tokens', 'cos_before_layer_og_vs_method_std', 'cos_delta_layer_og_vs_method_median', 'target_layer_bits', 'cos_after_mha_og_vs_method_num_tokens', 'mlp_in_affected_blocks_dense_bits', 'diff_norm_after_mha_og_vs_method_mean', 'norm_ratio_delta_mha_method_over_og_mean', 'norm_delta_mlp_to_layer_method_median', 'norm_after_mha_method_mean', 'norm_before_layer_method_std', 'rel_diff_norm_after_mha_residual_og_vs_method_num_tokens', 'norm_after_mha_og_mean', 'cos_before_layer_og_vs_method_num_tokens', 'cos_before_layer_og_vs_method_median', 'norm_delta_mlp_to_layer_method_mean', 'diff_norm_after_mlp_og_vs_method_std', 'overall_no_embeddings_dense_bits', 'diff_norm_after_mha_og_vs_method_num_tokens', 'lm_eval_attempted', 'norm_ratio_delta_mha_method_over_og_median', 'torch_compile_skip_reason', 'cos_after_mha_residual_og_vs_method_num_tokens', 'rel_diff_norm_after_layer_og_vs_method_num_tokens', 'rel_diff_norm_after_layer_og_vs_method_mean', 'norm_delta_mha_method_std', 'resume_key', 'superweight_abs_error_mean', 'cos_after_layer_og_vs_method_num_tokens', 'angle_deg_delta_mha_og_vs_method_median', 'norm_ratio_before_layer_method_over_og_mean', 'norm_ratio_delta_mlp_to_layer_method_over_og_num_tokens', 'diff_norm_after_mlp_og_vs_method_num_tokens', 'norm_after_mha_residual_og_std', 'affected_modules_storage_bits', 'cos_after_mlp_og_vs_method_std', 'diff_norm_after_mha_og_vs_method_std'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8/openai_gpt-oss-20b__partial_results__status.csv (at revision 5ea6bfa03283e398864b7526ef72b8ffa083895b), ['hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8@5ea6bfa03283e398864b7526ef72b8ffa083895b/gpt_oss_20b_mobe_b8_results.csv', 'hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8@5ea6bfa03283e398864b7526ef72b8ffa083895b/openai_gpt-oss-20b__partial_results.csv', 'hf://datasets/AverageMetaheuristicsEnjoyer/mobe-factor-cache-gpt-oss-20b-b8@5ea6bfa03283e398864b7526ef72b8ffa083895b/openai_gpt-oss-20b__partial_results__status.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

target_set string	model_arch_key string	model_family string	model_name string	method_label string	rank float64	target_modules string	target_layers string	storage_records string	affected_modules_dense_bits int64	affected_modules_storage_bits int64	affected_modules_compression_ratio float64	mha_in_affected_blocks_dense_bits float64	mha_in_affected_blocks_storage_bits float64	mha_in_affected_blocks_compression_ratio null	mlp_in_affected_blocks_dense_bits float64	mlp_in_affected_blocks_storage_bits float64	mlp_in_affected_blocks_compression_ratio null	affected_blocks_dense_bits int64	affected_blocks_storage_bits int64	affected_blocks_compression_ratio float64	overall_no_embeddings_dense_bits int64	overall_no_embeddings_storage_bits int64	overall_no_embeddings_compression_ratio float64	overall_with_embeddings_dense_bits int64	overall_with_embeddings_storage_bits int64	overall_with_embeddings_compression_ratio float64	target_layer_bits int64	overall_model_bits int64	target_layer_compression_ratio float64	overall_model_compression_ratio float64	superweight_abs_error_mean null	superweight_abs_error_max null	superweight_rel_error_mean null	superweight_rel_error_max null	superweight_error_detail string	target_layer_fro_rel_error float64	target_layer_max_abs_error float64	target_layer_error_detail string	mean_top_outlier_rel_error float64	mean_random_rel_error float64	decomposition_apply_time_s float64	quantization_apply_time_s float64	baseline_benchmarks_cached_per_model bool	ppl_requested bool	lm_eval_requested bool	hotpot_requested bool	generation_examples_requested bool	torch_compile_requested bool	torch_compile_applied bool	torch_compile_backend null	torch_compile_mode null	torch_compile_error null	torch_compile_error_traceback null	torch_compile_skip_reason null	torch_compile_device_summary string	ppl_wikitext2 float64	ppl_c4 float64	wikitext2_ppl float64	lm_eval_attempted bool	lm_eval_tasks_configured string	lm_eval_skipped_reason string	generation_examples string	generation_preview_prompt string	generation_preview_text string	resume_key string	moe_experts_in_affected_blocks_dense_bits float64	moe_experts_in_affected_blocks_storage_bits float64	moe_experts_in_affected_blocks_compression_ratio float64	recipe string	plan_stats string	benchmarked_with_dense_reconstruction bool	dense_reconstruction_records string	activation_geometry_stats string	activation_geometry_per_layer_stats string	cos_before_layer_og_vs_method_mean float64	cos_before_layer_og_vs_method_median float64	cos_before_layer_og_vs_method_std float64	cos_before_layer_og_vs_method_num_tokens float64	angle_deg_before_layer_og_vs_method_mean float64	angle_deg_before_layer_og_vs_method_median float64	angle_deg_before_layer_og_vs_method_std float64	angle_deg_before_layer_og_vs_method_num_tokens float64	norm_before_layer_og_mean float64	norm_before_layer_og_median float64	norm_before_layer_og_std float64	norm_before_layer_og_num_tokens float64	norm_before_layer_method_mean float64	norm_before_layer_method_median float64	norm_before_layer_method_std float64	norm_before_layer_method_num_tokens float64	norm_ratio_before_layer_method_over_og_mean float64	norm_ratio_before_layer_method_over_og_median float64	norm_ratio_before_layer_method_over_og_std float64	norm_ratio_before_layer_method_over_og_num_tokens float64	diff_norm_before_layer_og_vs_method_mean float64	diff_norm_before_layer_og_vs_method_median float64	diff_norm_before_layer_og_vs_method_std float64	diff_norm_before_layer_og_vs_method_num_tokens float64	rel_diff_norm_before_layer_og_vs_method_mean float64	rel_diff_norm_before_layer_og_vs_method_median float64	rel_diff_norm_before_layer_og_vs_method_std float64	rel_diff_norm_before_layer_og_vs_method_num_tokens float64	cos_after_mha_og_vs_method_mean float64	cos_after_mha_og_vs_method_median float64	cos_after_mha_og_vs_method_std float64	cos_after_mha_og_vs_method_num_tokens float64	angle_deg_after_mha_og_vs_method_mean float64	angle_deg_after_mha_og_vs_method_median float64	angle_deg_after_mha_og_vs_method_std float64	angle_deg_after_mha_og_vs_method_num_tokens float64	norm_after_mha_og_mean float64	norm_after_mha_og_median float64	norm_after_mha_og_std float64	norm_after_mha_og_num_tokens float64	norm_after_mha_method_mean float64	norm_after_mha_method_median float64	norm_after_mha_method_std float64	norm_after_mha_method_num_tokens float64	norm_ratio_after_mha_method_over_og_mean float64	norm_ratio_after_mha_method_over_og_median float64	norm_ratio_after_mha_method_over_og_std float64	norm_ratio_after_mha_method_over_og_num_tokens float64	diff_norm_after_mha_og_vs_method_mean float64	diff_norm_after_mha_og_vs_method_median float64	diff_norm_after_mha_og_vs_method_std float64	diff_norm_after_mha_og_vs_method_num_tokens float64	rel_diff_norm_after_mha_og_vs_method_mean float64	rel_diff_norm_after_mha_og_vs_method_median float64	rel_diff_norm_after_mha_og_vs_method_std float64	rel_diff_norm_after_mha_og_vs_method_num_tokens float64	cos_after_mha_residual_og_vs_method_mean float64	cos_after_mha_residual_og_vs_method_median float64	cos_after_mha_residual_og_vs_method_std float64	cos_after_mha_residual_og_vs_method_num_tokens float64	angle_deg_after_mha_residual_og_vs_method_mean float64	angle_deg_after_mha_residual_og_vs_method_median float64	angle_deg_after_mha_residual_og_vs_method_std float64	angle_deg_after_mha_residual_og_vs_method_num_tokens float64	norm_after_mha_residual_og_mean float64	norm_after_mha_residual_og_median float64	norm_after_mha_residual_og_std float64	norm_after_mha_residual_og_num_tokens float64	norm_after_mha_residual_method_mean float64	norm_after_mha_residual_method_median float64	norm_after_mha_residual_method_std float64	norm_after_mha_residual_method_num_tokens float64	norm_ratio_after_mha_residual_method_over_og_mean float64	norm_ratio_after_mha_residual_method_over_og_median float64	norm_ratio_after_mha_residual_method_over_og_std float64	norm_ratio_after_mha_residual_method_over_og_num_tokens float64	diff_norm_after_mha_residual_og_vs_method_mean float64	diff_norm_after_mha_residual_og_vs_method_median float64	diff_norm_after_mha_residual_og_vs_method_std float64	diff_norm_after_mha_residual_og_vs_method_num_tokens float64	rel_diff_norm_after_mha_residual_og_vs_method_mean float64	rel_diff_norm_after_mha_residual_og_vs_method_median float64	rel_diff_norm_after_mha_residual_og_vs_method_std float64	rel_diff_norm_after_mha_residual_og_vs_method_num_tokens float64	cos_after_mlp_og_vs_method_mean float64	cos_after_mlp_og_vs_method_median float64	cos_after_mlp_og_vs_method_std float64	cos_after_mlp_og_vs_method_num_tokens float64	angle_deg_after_mlp_og_vs_method_mean float64	angle_deg_after_mlp_og_vs_method_median float64	angle_deg_after_mlp_og_vs_method_std float64	angle_deg_after_mlp_og_vs_method_num_tokens float64	norm_after_mlp_og_mean float64	norm_after_mlp_og_median float64	norm_after_mlp_og_std float64	norm_after_mlp_og_num_tokens float64	norm_after_mlp_method_mean float64	norm_after_mlp_method_median float64	norm_after_mlp_method_std float64	norm_after_mlp_method_num_tokens float64	norm_ratio_after_mlp_method_over_og_mean float64	norm_ratio_after_mlp_method_over_og_median float64	norm_ratio_after_mlp_method_over_og_std float64	norm_ratio_after_mlp_method_over_og_num_tokens float64	diff_norm_after_mlp_og_vs_method_mean float64	diff_norm_after_mlp_og_vs_method_median float64	diff_norm_after_mlp_og_vs_method_std float64	diff_norm_after_mlp_og_vs_method_num_tokens float64	rel_diff_norm_after_mlp_og_vs_method_mean float64	rel_diff_norm_after_mlp_og_vs_method_median float64	rel_diff_norm_after_mlp_og_vs_method_std float64	rel_diff_norm_after_mlp_og_vs_method_num_tokens float64	cos_after_layer_og_vs_method_mean float64	cos_after_layer_og_vs_method_median float64	cos_after_layer_og_vs_method_std float64	cos_after_layer_og_vs_method_num_tokens float64	angle_deg_after_layer_og_vs_method_mean float64	angle_deg_after_layer_og_vs_method_median float64	angle_deg_after_layer_og_vs_method_std float64	angle_deg_after_layer_og_vs_method_num_tokens float64	norm_after_layer_og_mean float64	norm_after_layer_og_median float64	norm_after_layer_og_std float64	norm_after_layer_og_num_tokens float64	norm_after_layer_method_mean float64	norm_after_layer_method_median float64	norm_after_layer_method_std float64	norm_after_layer_method_num_tokens float64	norm_ratio_after_layer_method_over_og_mean float64	norm_ratio_after_layer_method_over_og_median float64	norm_ratio_after_layer_method_over_og_std float64	norm_ratio_after_layer_method_over_og_num_tokens float64	diff_norm_after_layer_og_vs_method_mean float64	diff_norm_after_layer_og_vs_method_median float64	diff_norm_after_layer_og_vs_method_std float64	diff_norm_after_layer_og_vs_method_num_tokens float64	rel_diff_norm_after_layer_og_vs_method_mean float64	rel_diff_norm_after_layer_og_vs_method_median float64	rel_diff_norm_after_layer_og_vs_method_std float64	rel_diff_norm_after_layer_og_vs_method_num_tokens float64	cos_delta_layer_og_vs_method_mean float64	cos_delta_layer_og_vs_method_median float64	cos_delta_layer_og_vs_method_std float64	cos_delta_layer_og_vs_method_num_tokens float64	angle_deg_delta_layer_og_vs_method_mean float64	angle_deg_delta_layer_og_vs_method_median float64	angle_deg_delta_layer_og_vs_method_std float64	angle_deg_delta_layer_og_vs_method_num_tokens float64	norm_delta_layer_og_mean float64	norm_delta_layer_og_median float64	norm_delta_layer_og_std float64	norm_delta_layer_og_num_tokens float64	norm_delta_layer_method_mean float64	norm_delta_layer_method_median float64	norm_delta_layer_method_std float64	norm_delta_layer_method_num_tokens float64	norm_ratio_delta_layer_method_over_og_mean float64	norm_ratio_delta_layer_method_over_og_median float64	norm_ratio_delta_layer_method_over_og_std float64	norm_ratio_delta_layer_method_over_og_num_tokens float64	diff_norm_delta_layer_og_vs_method_mean float64	diff_norm_delta_layer_og_vs_method_median float64	diff_norm_delta_layer_og_vs_method_std float64	diff_norm_delta_layer_og_vs_method_num_tokens float64	rel_diff_norm_delta_layer_og_vs_method_mean float64	rel_diff_norm_delta_layer_og_vs_method_median float64	rel_diff_norm_delta_layer_og_vs_method_std float64	rel_diff_norm_delta_layer_og_vs_method_num_tokens float64	cos_delta_mha_og_vs_method_mean float64	cos_delta_mha_og_vs_method_median float64	cos_delta_mha_og_vs_method_std float64	cos_delta_mha_og_vs_method_num_tokens float64	angle_deg_delta_mha_og_vs_method_mean float64	angle_deg_delta_mha_og_vs_method_median float64	angle_deg_delta_mha_og_vs_method_std float64	angle_deg_delta_mha_og_vs_method_num_tokens float64	norm_delta_mha_og_mean float64	norm_delta_mha_og_median float64	norm_delta_mha_og_std float64	norm_delta_mha_og_num_tokens float64	norm_delta_mha_method_mean float64	norm_delta_mha_method_median float64	norm_delta_mha_method_std float64	norm_delta_mha_method_num_tokens float64	norm_ratio_delta_mha_method_over_og_mean float64	norm_ratio_delta_mha_method_over_og_median float64	norm_ratio_delta_mha_method_over_og_std float64	norm_ratio_delta_mha_method_over_og_num_tokens float64	diff_norm_delta_mha_og_vs_method_mean float64	diff_norm_delta_mha_og_vs_method_median float64	diff_norm_delta_mha_og_vs_method_std float64	diff_norm_delta_mha_og_vs_method_num_tokens float64	rel_diff_norm_delta_mha_og_vs_method_mean float64	rel_diff_norm_delta_mha_og_vs_method_median float64	rel_diff_norm_delta_mha_og_vs_method_std float64	rel_diff_norm_delta_mha_og_vs_method_num_tokens float64	cos_delta_mlp_to_layer_og_vs_method_mean float64	cos_delta_mlp_to_layer_og_vs_method_median float64	cos_delta_mlp_to_layer_og_vs_method_std float64	cos_delta_mlp_to_layer_og_vs_method_num_tokens float64	angle_deg_delta_mlp_to_layer_og_vs_method_mean float64	angle_deg_delta_mlp_to_layer_og_vs_method_median float64	angle_deg_delta_mlp_to_layer_og_vs_method_std float64	angle_deg_delta_mlp_to_layer_og_vs_method_num_tokens float64	norm_delta_mlp_to_layer_og_mean float64	norm_delta_mlp_to_layer_og_median float64	norm_delta_mlp_to_layer_og_std float64	norm_delta_mlp_to_layer_og_num_tokens float64	norm_delta_mlp_to_layer_method_mean float64	norm_delta_mlp_to_layer_method_median float64	norm_delta_mlp_to_layer_method_std float64	norm_delta_mlp_to_layer_method_num_tokens float64	norm_ratio_delta_mlp_to_layer_method_over_og_mean float64	norm_ratio_delta_mlp_to_layer_method_over_og_median float64	norm_ratio_delta_mlp_to_layer_method_over_og_std float64	norm_ratio_delta_mlp_to_layer_method_over_og_num_tokens float64	diff_norm_delta_mlp_to_layer_og_vs_method_mean float64	diff_norm_delta_mlp_to_layer_og_vs_method_median float64	diff_norm_delta_mlp_to_layer_og_vs_method_std float64	diff_norm_delta_mlp_to_layer_og_vs_method_num_tokens float64	rel_diff_norm_delta_mlp_to_layer_og_vs_method_mean float64	rel_diff_norm_delta_mlp_to_layer_og_vs_method_median float64	rel_diff_norm_delta_mlp_to_layer_og_vs_method_std float64	rel_diff_norm_delta_mlp_to_layer_og_vs_method_num_tokens float64
__model_baseline__	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	baseline	null	[]	[]	[]	0	0	null	0	0	null	0	0	null	0	0	null	325,369,979,904	325,369,979,904	1	334,636,114,944	334,636,114,944	1	0	334,636,114,944	null	1	null	null	null	null	[]	null	null	[]	null	null	0	0	false	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	503.871613	269.941589	503.871613	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of a(...TRUNCATED)	['openai/gpt-oss-20b', '__model_baseline__', 'baseline']	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null
gpt_oss_20b_mobe_middle_out_layers_12_12	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	['model.layers.12.mlp.experts']	[12]	"[{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	8,493,465,600	5,094,612,992	1.667146	null	null	null	null	null	null	13,170,991,616	9,772,139,008	1.347811	325,369,979,904	321,971,127,296	1.010556	334,636,114,944	331,237,262,336	1.010261	5,094,612,992	331,237,262,336	1.667146	1.010261	null	null	null	null	[]	0.327671	0.388672	"[{'module': 'model.layers.12.mlp.experts', 'fro_rel_error': 0.32767105247557665, 'max_abs_error': 0(...TRUNCATED)	0.056051	0.467599	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	403.248779	224.596771	403.248779	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of a(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_12_12', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	8,493,465,600	5,094,612,992	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.12.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, 'std':(...TRUNCATED)	"{12: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	1.000001	1.000001	0	102,400	0.000007	0	0.000517	102,400	1,344.241577	1,305.57666	1,248.390991	102,400	1,344.241577	1,305.57666	1,248.390991	102,400	1	1	0	102,400	0	0	0	102,400	0	0	0	102,400	1	1	0	102,400	0.00111	0	0.006082	102,400	300.829041	288.386688	77.629799	102,400	300.829041	288.386688	77.629799	102,400	1	1	0	102,400	0	0	0	102,400	0	0	0	102,400	1.000001	1.000001	0	102,400	0.000002	0	0.000321	102,400	1,277.930786	1,249.170166	1,250.1604	102,400	1,277.930786	1,249.170166	1,250.1604	102,400	1	1	0	102,400	0	0	0	102,400	0	0	0	102,400	0.818728	0.85094	0.117244	102,400	33.571514	31.685959	10.93597	102,400	560.145996	544.13092	190.614685	102,400	332.689056	313.372711	99.031631	102,400	0.616296	0.595402	0.139274	102,400	333.134247	323.896057	105.684456	102,400	0.614229	0.602364	0.12466	102,400	0.984029	0.9852	0.006395	102,400	10.066194	9.869831	1.963596	102,400	1,627.13916	1,588.779297	1,248.423096	102,400	1,423.192993	1,386.245483	1,247.569458	102,400	0.874306	0.876913	0.040214	102,400	333.156799	323.937622	105.676201	102,400	0.207141	0.203358	0.044833	102,400	0.727228	0.751938	0.13249	102,400	42.265411	41.241493	10.8617	102,400	490.230713	472.366577	132.352768	102,400	377.687958	370.904175	64.624016	102,400	0.794985	0.787387	0.153119	102,400	333.156799	323.937622	105.676201	102,400	0.684575	0.676639	0.144178	102,400	1	1	0	102,400	0.000009	0	0.000463	102,400	300.837708	288.368835	77.630577	102,400	300.837708	288.368835	77.630577	102,400	1	1	0	102,400	0	0	0	102,400	0	0	0	102,400	0.818695	0.850863	0.11725	102,400	33.575062	31.694336	10.935762	102,400	560.151917	544.191772	190.606232	102,400	332.694336	313.335144	99.030014	102,400	0.616303	0.595346	0.139293	102,400	333.156799	323.937622	105.676201	102,400	0.61426	0.602406	0.124654	102,400
gpt_oss_20b_mobe_middle_out_layers_12_13	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	['model.layers.12.mlp.experts', 'model.layers.13.mlp.experts']	[12, 13]	"[{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	16,986,931,200	10,189,225,984	1.667146	null	null	null	null	null	null	26,341,983,232	19,544,278,016	1.347811	325,369,979,904	318,572,274,688	1.021338	334,636,114,944	327,838,409,728	1.020735	10,189,225,984	327,838,409,728	1.667146	1.020735	null	null	null	null	[]	0.326333	0.388672	"[{'module': 'model.layers.12.mlp.experts', 'fro_rel_error': 0.32767105247557665, 'max_abs_error': 0(...TRUNCATED)	0.032633	0.456765	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	352.073212	201.462448	352.073212	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of a(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_12_13', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	16,986,931,200	10,189,225,984	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.12.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.9920149445533752, 'median': 0.999998152256012, 'std': (...TRUNCATED)	"{12: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	0.992015	0.999998	0.009177	204,800	5.033101	0.104901	5.221111	204,800	1,485.690308	1,421.543945	1,256.391846	204,800	1,383.717163	1,344.438721	1,248.601318	204,800	0.937153	1	0.068981	204,800	166.5784	53.323021	182.571075	204,800	0.103571	0.001386	0.108314	204,800	0.774899	0.999978	0.291333	204,800	27.47422	0.303889	30.254776	204,800	349.424194	332.342621	110.408577	204,800	364.841705	363.378784	103.400314	204,800	1.07994	1	0.300712	204,800	195.657776	1.154692	210.970398	204,800	0.525144	0.004689	0.602114	204,800	0.992849	0.999998	0.008496	204,800	4.750605	0.092837	4.951705	204,800	1,397.004639	1,351.952759	1,252.901978	204,800	1,415.882446	1,364.832764	1,254.43396	204,800	1.013209	1	0.033233	204,800	127.822418	48.727715	134.691833	204,800	0.087059	0.001377	0.091536	204,800	0.831298	0.866906	0.120332	204,800	32.219135	29.898895	11.059153	204,800	596.307007	587.995972	172.380447	204,800	393.933136	370.465118	132.312561	204,800	0.666845	0.662811	0.137096	204,800	329.692566	320.287048	91.312134	204,800	0.574128	0.5612	0.140922	204,800	0.982237	0.983358	0.007069	204,800	10.618133	10.46752	2.069837	204,800	1,766.085938	1,734.776611	1,249.052246	204,800	1,598.792847	1,558.585693	1,254.177734	204,800	0.902623	0.906308	0.046317	204,800	354.722107	351.118286	92.86097	204,800	0.204468	0.201168	0.040918	204,800	0.790538	0.829976	0.121466	204,800	36.558952	33.903683	10.49609	204,800	533.329712	521.180115	129.167236	204,800	507.328247	495.740784	151.198029	204,800	0.958703	0.971878	0.223427	204,800	332.573059	323.580322	88.203247	204,800	0.637647	0.615439	0.151689	204,800	0.774888	0.999949	0.291338	204,800	27.474783	0.425798	30.255749	204,800	349.433136	332.331299	110.409798	204,800	364.848083	363.375244	103.398643	204,800	1.079931	1	0.300693	204,800	195.666794	1.786366	210.976318	204,800	0.525156	0.007204	0.602111	204,800	0.831268	0.866863	0.120331	204,800	32.222588	29.903919	11.058534	204,800	596.310425	588.029907	172.375443	204,800	393.938995	370.472656	132.311554	204,800	0.666855	0.662878	0.137112	204,800	329.716339	320.32016	91.30677	204,800	0.574163	0.561245	0.140913	204,800
gpt_oss_20b_mobe_middle_out_layers_12_14	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	['model.layers.12.mlp.experts', 'model.layers.13.mlp.experts', 'model.layers.14.mlp.experts']	[12, 13, 14]	"[{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	25,480,396,800	15,283,838,976	1.667146	null	null	null	null	null	null	39,512,974,848	29,316,417,024	1.347811	325,369,979,904	315,173,422,080	1.032352	334,636,114,944	324,439,557,120	1.031428	15,283,838,976	324,439,557,120	1.667146	1.031428	null	null	null	null	[]	0.324867	0.414063	"[{'module': 'model.layers.12.mlp.experts', 'fro_rel_error': 0.32767105247557665, 'max_abs_error': 0(...TRUNCATED)	0.030524	0.45707	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	404.872223	205.147751	404.872223	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of a(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_12_14', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	25,480,396,800	15,283,838,976	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.12.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.9881580471992493, 'median': 0.9872336983680725, 'std':(...TRUNCATED)	"{12: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	0.988158	0.987234	0.01017	307,200	7.078757	9.165018	5.283046	307,200	1,625.471191	1,581.55835	1,264.563599	307,200	1,513.942261	1,428.804443	1,257.985962	307,200	0.935082	0.935068	0.059476	307,200	236.4814	297.477905	183.604401	307,200	0.136312	0.176548	0.102013	307,200	0.797427	0.863164	0.243796	307,200	28.920494	30.326315	25.172091	307,200	377.537384	374.607361	104.841316	307,200	404.044769	410.657135	109.281143	307,200	1.092838	1	0.257739	307,200	216.91983	247.90686	179.015121	307,200	0.550837	0.567017	0.500715	307,200	0.989193	0.988351	0.00952	307,200	6.750803	8.753963	5.06065	307,200	1,572.967407	1,483.82373	1,270.757446	307,200	1,584.216431	1,515.668823	1,269.760376	307,200	1.008184	1	0.031418	307,200	203.279724	246.867645	156.857513	307,200	0.120896	0.156545	0.090894	307,200	0.836512	0.863504	0.105896	307,200	31.953262	30.287668	9.918629	307,200	691.107666	667.302734	221.688232	307,200	500.870422	465.038208	204.963913	307,200	0.718276	0.729496	0.180148	307,200	372.532013	365.562805	109.216171	307,200	0.561923	0.54582	0.166653	307,200	0.979284	0.980642	0.008965	307,200	11.4346	11.291927	2.413607	307,200	2,003.269165	1,880.019043	1,287.41272	307,200	1,818.85144	1,737.07251	1,285.724976	307,200	0.905411	0.909556	0.041436	307,200	427.987396	404.145752	140.282211	307,200	0.215001	0.212919	0.042398	307,200	0.826258	0.864357	0.113344	307,200	32.963299	30.190632	10.309073	307,200	664.235535	596.437683	227.035263	307,200	639.086121	630.496216	236.773666	307,200	0.966405	0.972256	0.215827	307,200	358.842072	352.813324	93.678589	307,200	0.576278	0.543626	0.188359	307,200	0.797403	0.863136	0.243797	307,200	28.923542	30.329411	25.171959	307,200	377.546814	374.596191	104.843163	307,200	404.051361	410.670105	109.279778	307,200	1.09283	1	0.257724	307,200	216.940491	247.890656	179.021606	307,200	0.550889	0.567111	0.500697	307,200	0.836484	0.863413	0.105896	307,200	31.956442	30.298038	9.918156	307,200	691.110779	667.349243	221.687134	307,200	500.875305	465.080322	204.962601	307,200	0.718286	0.729551	0.180214	307,200	372.559052	365.620056	109.218063	307,200	0.56196	0.545843	0.166706	307,200
gpt_oss_20b_mobe_middle_out_layers_12_15	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	"['model.layers.12.mlp.experts', 'model.layers.13.mlp.experts', 'model.layers.14.mlp.experts', 'mode(...TRUNCATED)	[12, 13, 14, 15]	"[{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	33,973,862,400	20,378,451,968	1.667146	null	null	null	null	null	null	52,683,966,464	39,088,556,032	1.347811	325,369,979,904	311,774,569,472	1.043607	334,636,114,944	321,040,704,512	1.042348	20,378,451,968	321,040,704,512	1.667146	1.042348	null	null	null	null	[]	0.32424	0.414063	"[{'module': 'model.layers.12.mlp.experts', 'fro_rel_error': 0.32767105247557665, 'max_abs_error': 0(...TRUNCATED)	0.023387	0.469599	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	404.330292	205.147751	404.330292	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of a(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_12_15', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	33,973,862,400	20,378,451,968	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.12.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.984463095664978, 'median': 0.983917772769928, 'std': 0(...TRUNCATED)	"{12: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	0.984463	0.983918	0.011864	409,600	8.575953	10.289505	5.374457	409,600	1,838.512329	1,734.224609	1,309.245728	409,600	1,700.198853	1,564.153564	1,292.929932	409,600	0.929058	0.924213	0.054454	409,600	320.99057	351.032135	221.595078	409,600	0.161251	0.196131	0.100077	409,600	0.76967	0.815924	0.227583	409,600	33.103378	35.321167	23.626623	409,600	424.528442	415.595032	146.230972	409,600	496.53949	454.958313	205.724411	409,600	1.172576	1.073864	0.31621	409,600	304.188568	306.047302	230.264389	409,600	0.673202	0.683932	0.512311	409,600	0.98422	0.985034	0.013169	409,600	8.570705	9.924916	5.533073	409,600	1,794.092651	1,657.943848	1,317.379883	409,600	1,830.290894	1,681.794189	1,332.562622	409,600	1.017287	1.003054	0.038858	409,600	307.355042	298.752899	231.881332	409,600	0.15496	0.177324	0.10096	409,600	0.812785	0.842126	0.117658	409,600	34.257248	32.634678	10.731263	409,600	758.87439	743.864929	243.536896	409,600	554.055298	544.637939	219.257156	409,600	0.724646	0.730452	0.166945	409,600	436.307983	411.767151	156.371124	409,600	0.587911	0.573262	0.164633	409,600	0.975232	0.977792	0.012659	409,600	12.416644	12.097501	3.049855	409,600	2,260.985107	2,082.342285	1,346.817871	409,600	2,072.241943	1,908.979736	1,345.216553	409,600	0.912191	0.915658	0.04122	409,600	520.8797	482.583466	210.469894	409,600	0.228337	0.224261	0.049359	409,600	0.835489	0.866148	0.101584	409,600	32.226181	29.985975	9.302603	409,600	766.412109	739.088623	277.237915	409,600	754.974243	742.026001	298.228821	409,600	0.982935	0.991878	0.194681	409,600	411.150269	386.605225	131.442383	409,600	0.565638	0.536397	0.169563	409,600	0.769644	0.815906	0.227587	409,600	33.106251	35.322983	23.626514	409,600	424.537994	415.620605	146.231781	409,600	496.545227	454.986938	205.722702	409,600	1.172564	1.073885	0.316193	409,600	304.209869	306.089569	230.271255	409,600	0.673246	0.683992	0.512293	409,600	0.812756	0.842114	0.11766	409,600	34.260319	32.635963	10.730865	409,600	758.875305	743.829102	243.53447	409,600	554.06073	544.575806	219.256439	409,600	0.724657	0.730412	0.166999	409,600	436.335236	411.851196	156.373093	409,600	0.587947	0.57332	0.164673	409,600
gpt_oss_20b_mobe_middle_out_layers_12_16	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	"['model.layers.12.mlp.experts', 'model.layers.13.mlp.experts', 'model.layers.14.mlp.experts', 'mode(...TRUNCATED)	[12, 13, 14, 15, 16]	"[{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	42,467,328,000	25,473,064,960	1.667146	null	null	null	null	null	null	65,854,958,080	48,860,695,040	1.347811	325,369,979,904	308,375,716,864	1.055109	334,636,114,944	317,641,851,904	1.053501	25,473,064,960	317,641,851,904	1.667146	1.053501	null	null	null	null	[]	0.325553	0.414063	"[{'module': 'model.layers.12.mlp.experts', 'fro_rel_error': 0.32767105247557665, 'max_abs_error': 0(...TRUNCATED)	0.019214	0.468599	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	475.8862	230.316193	475.8862	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of a(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_12_16', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	42,467,328,000	25,473,064,960	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.12.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.9801855087280273, 'median': 0.9808904528617859, 'std':(...TRUNCATED)	"{12: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	0.980186	0.98089	0.015045	512,000	9.933316	11.21907	5.66648	512,000	2,077.636475	1,880.317871	1,377.423218	512,000	1,926.641724	1,737.411865	1,358.004639	512,000	0.929753	0.926339	0.050921	512,000	416.703735	404.119263	280.799988	512,000	0.18267	0.210468	0.101445	512,000	0.784993	0.832502	0.208073	512,000	32.78231	33.643417	21.376999	512,000	484.837067	454.861633	185.238358	512,000	581.253723	504.069031	263.692261	512,000	1.191701	1.133513	0.295036	512,000	341.232819	352.003845	225.629547	512,000	0.674694	0.678099	0.464257	512,000	0.981714	0.982056	0.013671	512,000	9.549046	10.87044	5.426795	512,000	2,085.872314	1,889.704102	1,423.619629	512,000	2,126.594727	1,893.613281	1,439.474854	512,000	1.017483	1.007311	0.0375	512,000	398.675751	354.40332	280.821228	512,000	0.171787	0.193934	0.098162	512,000	0.807808	0.834985	0.115869	512,000	34.796375	33.385727	10.592464	512,000	875.108215	832.12854	344.380554	512,000	655.890869	627.484192	297.721161	512,000	0.740536	0.747689	0.160659	512,000	509.720123	462.098816	215.311752	512,000	0.593739	0.578883	0.160344	512,000	0.972774	0.975476	0.013917	512,000	13.011203	12.715226	3.240585	512,000	2,639.976074	2,442.887695	1,526.758179	512,000	2,425.270264	2,227.255371	1,502.462036	512,000	0.91434	0.917977	0.040334	512,000	638.256104	568.282227	311.293549	512,000	0.236803	0.232548	0.05177	512,000	0.849443	0.876511	0.097021	512,000	30.698484	28.775721	9.194313	512,000	941.85791	895.368652	444.132507	512,000	923.411621	885.09375	441.922424	512,000	0.981461	0.988083	0.178127	512,000	467.864075	426.738617	174.676682	512,000	0.538331	0.511648	0.166366	512,000	0.784962	0.832428	0.208077	512,000	32.785721	33.651073	21.376837	512,000	484.847748	454.86319	185.240005	512,000	581.260376	504.074554	263.692474	512,000	1.191688	1.133531	0.295021	512,000	341.263519	351.995026	225.642151	512,000	0.674746	0.678201	0.464241	512,000	0.807777	0.834962	0.115873	512,000	34.799587	33.388153	10.592143	512,000	875.111328	832.12439	344.383575	512,000	655.898499	627.505981	297.728485	512,000	0.740545	0.747714	0.160705	512,000	509.759491	462.137207	215.34108	512,000	0.593778	0.578934	0.160377	512,000
gpt_oss_20b_mobe_middle_out_layers_12_17	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	"['model.layers.12.mlp.experts', 'model.layers.13.mlp.experts', 'model.layers.14.mlp.experts', 'mode(...TRUNCATED)	[12, 13, 14, 15, 16, 17]	"[{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	50,960,793,600	30,567,677,952	1.667146	null	null	null	null	null	null	79,025,949,696	58,632,834,048	1.347811	325,369,979,904	304,976,864,256	1.066868	334,636,114,944	314,242,999,296	1.064896	30,567,677,952	314,242,999,296	1.667146	1.064896	null	null	null	null	[]	0.324587	0.414063	"[{'module': 'model.layers.12.mlp.experts', 'fro_rel_error': 0.32767105247557665, 'max_abs_error': 0(...TRUNCATED)	0.016086	0.47422	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	389.970703	207.705368	389.970703	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of t(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_12_17', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	50,960,793,600	30,567,677,952	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.12.mlp.experts', 'layer_idx': 12, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.12.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.977311372756958, 'median': 0.9783252477645874, 'std': (...TRUNCATED)	"{12: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	0.977311	0.978325	0.016259	614,400	10.84267	11.950954	5.680129	614,400	2,424.020508	2,082.993652	1,560.583618	614,400	2,245.098633	1,909.573975	1,517.634155	614,400	0.928617	0.926234	0.048732	614,400	531.880127	482.583466	370.583893	614,400	0.197336	0.221028	0.100109	614,400	0.753021	0.805364	0.221519	614,400	36.013393	36.354603	21.730356	614,400	560.583862	498.597229	250.892471	614,400	721.253479	586.479492	419.393616	614,400	1.251459	1.18127	0.346589	614,400	473.713867	406.750305	383.689331	614,400	0.771261	0.744476	0.508626	614,400	0.978975	0.979473	0.015094	614,400	10.43659	11.629042	5.464517	614,400	2,428.504639	2,069.985352	1,588.734375	614,400	2,536.208252	2,073.299805	1,683.768433	614,400	1.032359	1.016305	0.054218	614,400	534.61792	462.122955	409.021179	614,400	0.192741	0.209173	0.104298	614,400	0.809697	0.835807	0.113167	614,400	34.648342	33.300007	10.406281	614,400	1,025.315186	919.604309	685.414978	614,400	748.797058	702.201782	498.732758	614,400	0.730066	0.731851	0.153163	614,400	591.39801	526.20459	315.247009	614,400	0.591575	0.577844	0.153798	614,400	0.971752	0.974346	0.015143	614,400	13.264176	13.006208	3.266675	614,400	3,080.009277	2,687.3125	1,716.786865	614,400	2,878.47583	2,456.358398	1,722.382813	614,400	0.924363	0.924576	0.055629	614,400	754.327026	668.763672	419.724091	614,400	0.239956	0.235172	0.062364	614,400	0.855513	0.879972	0.091818	614,400	30.097084	28.361073	8.819551	614,400	1,124.123291	992.535767	761.697144	614,400	1,115.468384	995.57959	683.186523	614,400	0.990045	0.994061	0.169637	614,400	551.44397	482.520752	297.460297	614,400	0.529917	0.504228	0.160344	614,400	0.752992	0.805286	0.221521	614,400	36.016548	36.362144	21.729994	614,400	560.596741	498.629333	250.897232	614,400	721.261536	586.470642	419.396149	614,400	1.251444	1.181331	0.346573	614,400	473.746765	406.771545	383.700104	614,400	0.771306	0.744493	0.508605	614,400	0.809666	0.835772	0.113169	614,400	34.651688	33.303677	10.40594	614,400	1,025.319824	919.672241	685.41803	614,400	748.806946	702.229126	498.737457	614,400	0.730075	0.731851	0.153203	614,400	591.443665	526.254272	315.272766	614,400	0.591615	0.577894	0.153827	614,400
gpt_oss_20b_mobe_middle_out_layers_11_17	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	"['model.layers.11.mlp.experts', 'model.layers.12.mlp.experts', 'model.layers.13.mlp.experts', 'mode(...TRUNCATED)	[11, 12, 13, 14, 15, 16, 17]	"[{'full_name': 'model.layers.11.mlp.experts', 'layer_idx': 11, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	59,454,259,200	35,662,290,944	1.667146	null	null	null	null	null	null	92,196,941,312	68,404,973,056	1.347811	325,369,979,904	301,578,011,648	1.078892	334,636,114,944	310,844,146,688	1.07654	35,662,290,944	310,844,146,688	1.667146	1.07654	null	null	null	null	[]	0.324244	0.414063	"[{'module': 'model.layers.11.mlp.experts', 'fro_rel_error': 0.3222554889733331, 'max_abs_error': 0.(...TRUNCATED)	0.016947	0.471392	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	355.389465	201.321091	355.389465	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of a(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_11_17', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	59,454,259,200	35,662,290,944	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.11.mlp.experts', 'layer_idx': 11, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.11.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.9765776991844177, 'median': 0.9772055745124817, 'std':(...TRUNCATED)	"{11: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	0.976578	0.977206	0.017976	716,800	10.979378	12.256893	5.848226	716,800	2,245.275635	1,880.783081	1,582.497925	716,800	2,076.415771	1,731.700439	1,538.423706	716,800	0.9273	0.927344	0.048928	716,800	507.179962	443.909058	377.940704	716,800	0.199798	0.225794	0.102299	716,800	0.757031	0.81639	0.229224	716,800	35.620487	35.274994	21.856518	716,800	522.276978	457.172882	252.000961	716,800	658.86084	506.733673	422.594727	716,800	1.212133	1.127942	0.357717	716,800	437.634521	370.843689	384.522858	716,800	0.750602	0.703051	0.511443	716,800	0.978106	0.978177	0.017302	716,800	10.549946	11.991884	5.771494	716,800	2,239.647217	1,890.061279	1,613.1875	716,800	2,332.621338	1,886.071777	1,710.028687	716,800	1.02662	1.006816	0.056585	716,800	508.644104	399.980652	420.208313	716,800	0.193972	0.215116	0.109427	716,800	0.81428	0.842	0.12136	716,800	33.903286	32.648064	11.396422	716,800	937.599487	832.087158	671.159058	716,800	683.318115	624.345215	489.286407	716,800	0.729332	0.72902	0.147527	716,800	541.664978	471.503296	329.08844	716,800	0.581043	0.568874	0.163086	716,800	0.971511	0.973399	0.017334	716,800	13.117727	13.245113	4.026303	716,800	2,832.042236	2,443.388184	1,765.748901	716,800	2,644.855957	2,225.633301	1,761.664429	716,800	0.923844	0.925919	0.054802	716,800	703.224976	603.50946	445.351105	716,800	0.237387	0.239671	0.072829	716,800	0.869003	0.880209	0.062023	716,800	28.976015	28.332462	6.69441	716,800	1,017.244446	895.421631	752.767822	716,800	1,006.630005	888.385681	688.383667	716,800	0.980969	0.9784	0.162745	716,800	501.962677	434.268188	317.205139	716,800	0.514464	0.498092	0.142451	716,800	0.757	0.816365	0.229222	716,800	35.624249	35.277496	21.85537	716,800	522.289124	457.180359	252.005875	716,800	658.868103	506.74707	422.593384	716,800	1.21212	1.128016	0.357699	716,800	437.666504	370.855225	384.527924	716,800	0.750659	0.703116	0.511407	716,800	0.814247	0.841954	0.121364	716,800	33.906937	32.652946	11.395921	716,800	937.603333	832.053955	671.162231	716,800	683.328186	624.345154	489.290619	716,800	0.729344	0.728965	0.147564	716,800	541.707336	471.494507	329.114746	716,800	0.581085	0.568937	0.163109	716,800
gpt_oss_20b_mobe_middle_out_layers_11_18	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	"['model.layers.11.mlp.experts', 'model.layers.12.mlp.experts', 'model.layers.13.mlp.experts', 'mode(...TRUNCATED)	[11, 12, 13, 14, 15, 16, 17, 18]	"[{'full_name': 'model.layers.11.mlp.experts', 'layer_idx': 11, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	67,947,724,800	40,756,903,936	1.667146	null	null	null	null	null	null	105,367,932,928	78,177,112,064	1.347811	325,369,979,904	298,179,159,040	1.09119	334,636,114,944	307,445,294,080	1.088441	40,756,903,936	307,445,294,080	1.667146	1.088441	null	null	null	null	[]	0.32312	0.414063	"[{'module': 'model.layers.11.mlp.experts', 'fro_rel_error': 0.3222554889733331, 'max_abs_error': 0.(...TRUNCATED)	0.014847	0.477685	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	426.962585	217.67337	426.962585	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of t(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_11_18', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	67,947,724,800	40,756,903,936	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.11.mlp.experts', 'layer_idx': 11, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.11.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.9750721454620361, 'median': 0.975639820098877, 'std': (...TRUNCATED)	"{11: {'cos_before_layer_og_vs_method': {'mean': 1.0000005960464478, 'median': 1.0000005960464478, '(...TRUNCATED)	0.975072	0.97564	0.018753	819,200	11.478014	12.672534	5.745031	819,200	2,624.638184	2,083.836914	1,796.166016	819,200	2,460.850098	1,904.25293	1,774.682251	819,200	0.933363	0.932397	0.057116	819,200	615.321899	519.305603	477.110199	819,200	0.207713	0.230608	0.103945	819,200	0.7745	0.837026	0.220692	819,200	34.309757	33.172607	20.928299	819,200	604.737915	499.798492	328.085236	819,200	751.780334	590.923706	474.70874	819,200	1.209326	1.140864	0.338696	819,200	458.627502	413.944489	369.116058	819,200	0.722405	0.660988	0.488029	819,200	0.977309	0.977632	0.017753	819,200	10.892666	12.141273	5.590516	819,200	2,682.379395	2,070.504395	1,920.338623	819,200	2,779.748535	2,066.822021	2,003.868164	819,200	1.026046	1.009656	0.058923	819,200	614.777466	510.555664	504.279022	819,200	0.199529	0.217921	0.108378	819,200	0.817621	0.843586	0.117522	819,200	33.642365	32.479282	11.081781	819,200	1,135.106445	919.617676	832.123657	819,200	851.476074	701.003296	667.923035	819,200	0.739328	0.740516	0.14997	819,200	643.629822	538.210083	449.116333	819,200	0.576731	0.563534	0.163765	819,200	0.970772	0.972994	0.01794	819,200	13.306	13.345922	4.020214	819,200	3,427.245361	2,687.933105	2,301.181641	819,200	3,205.166504	2,454.796875	2,235.308594	819,200	0.9257	0.928069	0.052866	819,200	854.892517	704.55835	591.075562	819,200	0.239759	0.240416	0.071661	819,200	0.874169	0.885335	0.063083	819,200	28.302912	27.707266	6.934573	819,200	1,278.024902	992.542603	995.133179	819,200	1,238.319946	999.678711	908.43689	819,200	0.973741	0.967752	0.158443	819,200	596.167603	495.451416	431.48526	819,200	0.501281	0.485407	0.147424	819,200	0.774466	0.837001	0.220689	819,200	34.314171	33.175247	20.92695	819,200	604.751709	499.840302	328.091187	819,200	751.789246	590.956421	474.709045	819,200	1.209314	1.140838	0.338679	819,200	458.677765	413.996918	369.12619	819,200	0.722473	0.661024	0.487991	819,200	0.817587	0.843556	0.117526	819,200	33.646114	32.482426	11.081266	819,200	1,135.11145	919.682007	832.128662	819,200	851.485718	701.015015	667.924072	819,200	0.739338	0.740524	0.150004	819,200	643.685852	538.261353	449.15506	819,200	0.576775	0.563634	0.163787	819,200
gpt_oss_20b_mobe_middle_out_layers_10_18	gpt_oss_20b	gpt_oss_moe	openai/gpt-oss-20b	mobe_gate_up_b8_t1382	0	"['model.layers.10.mlp.experts', 'model.layers.11.mlp.experts', 'model.layers.12.mlp.experts', 'mode(...TRUNCATED)	[10, 11, 12, 13, 14, 15, 16, 17, 18]	"[{'full_name': 'model.layers.10.mlp.experts', 'layer_idx': 10, 'module_name': 'mlp.experts', 'modul(...TRUNCATED)	76,441,190,400	45,851,516,928	1.667146	null	null	null	null	null	null	118,538,924,544	87,949,251,072	1.347811	325,369,979,904	294,780,306,432	1.103771	334,636,114,944	304,046,441,472	1.100609	45,851,516,928	304,046,441,472	1.667146	1.100609	null	null	null	null	[]	0.323587	0.414063	"[{'module': 'model.layers.10.mlp.experts', 'fro_rel_error': 0.32741438204382894, 'max_abs_error': 0(...TRUNCATED)	0.013552	0.478465	null	null	null	true	false	false	true	false	false	null	null	null	null	null	{'param_devices': ['cuda:0'], 'buffer_devices': ['cuda:0'], 'hf_device_map': None}	410.331116	212.254944	410.331116	false	['arc_challenge', 'winogrande', 'piqa', 'hellaswag', 'openbookqa']	cfg.run_lm_eval is False	"{'The theory of tensor train decomposition for neural networks suggests that': 'The theory of tenso(...TRUNCATED)	The theory of tensor train decomposition for neural networks suggests that	"The theory of tensor train decomposition for neural networks suggests that the weight matrices of t(...TRUNCATED)	"['openai/gpt-oss-20b', 'gpt_oss_20b_mobe_middle_out_layers_10_18', 'mobe_gate_up_b8_t1382', 0, '[{\(...TRUNCATED)	76,441,190,400	45,851,516,928	1.667146	"{'label': 'mobe_gate_up_b8_t1382', 'plans': [{'target': 'moe_experts', 'decomposition_method': 'mob(...TRUNCATED)	"[{'decomposition_per_module': [{'full_name': 'model.layers.10.mlp.experts', 'layer_idx': 10, 'modul(...TRUNCATED)	true	"[{'module': 'model.layers.10.mlp.experts', 'old_type': 'GptOssExperts', 'new_type': 'GptOssExperts'(...TRUNCATED)	"{'cos_before_layer_og_vs_method': {'mean': 0.9705455303192139, 'median': 0.9718649387359619, 'std':(...TRUNCATED)	"{10: {'cos_before_layer_og_vs_method': {'mean': 1.0000004768371582, 'median': 1.0000004768371582, '(...TRUNCATED)	0.970546	0.971865	0.021884	921,600	12.62078	13.623394	5.966991	921,600	2,440.499268	1,881.059326	1,820.73291	921,600	2,278.411621	1,722.19165	1,797.526123	921,600	0.92752	0.926451	0.058381	921,600	620.682129	501.483337	487.175385	921,600	0.22718	0.245794	0.106043	921,600	0.738224	0.796251	0.228764	921,600	37.967224	37.226429	20.99765	921,600	564.419434	457.183411	330.079926	921,600	697.327942	510.813141	475.304565	921,600	1.197953	1.131647	0.351242	921,600	453.965149	390.889771	360.019836	921,600	0.779943	0.723799	0.482096	921,600	0.972835	0.974302	0.021069	921,600	12.060822	13.017349	5.848926	921,600	2,488.251221	1,890.385742	1,937.97937	921,600	2,577.332031	1,881.720703	2,018.527222	921,600	1.025258	1.011669	0.05902	921,600	620.862793	469.367249	516.178162	921,600	0.219486	0.233322	0.111047	921,600	0.796527	0.82838	0.132147	921,600	35.628494	34.067268	11.792784	921,600	1,052.969482	832.854858	819.196655	921,600	779.869324	621.685791	661.340332	921,600	0.722498	0.718267	0.153641	921,600	617.237366	497.804382	444.301941	921,600	0.605476	0.587546	0.165314	921,600	0.966356	0.9693	0.02099	921,600	14.317059	14.233903	4.203971	921,600	3,176.752197	2,443.791016	2,320.367432	921,600	2,962.068359	2,220.676025	2,253.377441	921,600	0.920591	0.922853	0.054127	921,600	842.073547	664.239563	608.054626	921,600	0.256838	0.25497	0.07278	921,600	0.853143	0.869569	0.078452	921,600	30.536165	29.591454	8.020005	921,600	1,177.22583	896.15686	980.989502	921,600	1,143.53064	896.874146	897.116638	921,600	0.983713	0.979913	0.162865	921,600	579.116089	466.779877	431.055206	921,600	0.53962	0.521098	0.158104	921,600	0.73819	0.796226	0.228762	921,600	37.971119	37.228806	20.996492	921,600	564.43219	457.186523	330.086151	921,600	697.336731	510.829346	475.306702	921,600	1.197943	1.131642	0.351231	921,600	454.010345	390.996094	360.033875	921,600	0.780002	0.723837	0.482067	921,600	0.796493	0.828352	0.132152	921,600	35.632027	34.070133	11.792472	921,600	1,052.973999	832.922546	819.201477	921,600	779.880005	621.672485	661.346558	921,600	0.722508	0.718269	0.153669	921,600	617.283752	497.782593	444.332611	921,600	0.605517	0.587582	0.165333	921,600

End of preview.

MoBE factor cache — GPT-OSS-20B (num_B=8)

Trained MoBE (Mixture-of-Basis-Experts) factors for the MoE gate_proj / up_proj matrices of GPT-OSS-20B, plus the perplexity / activation-geometry results of the compression sweep that produced them.

mobe_factor_cache/<model>/layer_<i>/<projection>__nB<num_B>_t<T>_<activation>_E<E>_R<R>_C<C>.pt Each file is a torch.save payload with factors (A, B, W), target_std, meta (hyperparameters, final loss, best epoch, training time) and the loss history. Reconstruct with reconstruct_from_factors(A, B, W, activation, target_std).
*_results.csv / *_results.json — perplexity (WikiText-2, C4), compression ratios and residual-stream activation geometry for each cumulative set of compressed MoE layers.
*__partial_results* — per-target-set checkpoints written during the sweep.

Configuration: num_B=8, activation tanh, 10000 epochs per projection.

These are derived caches, not a standalone model.

Downloads last month: 36

Total file size:

15.3 GB

MoBE factor cache — GPT-OSS-20B (num_B=8)

Contents