steps:
- ekfac:
    index_cfg:
      run_path: /mnt/ssd-2/lucia/paper_runs/experiments/plan_adam_eps1e17_64k_bs256/ekfac_scores
      overwrite: true
      model: /mnt/ssd-2/lucia/paper_runs/experiments/plan_adam_eps1e17_64k_bs256/retrained/base
      precision: fp32
      revision: null
      distributed:
        nnode: 1
        nproc_per_node: 2
        node_rank: null
      fsdp: false
      peft_init_kwargs: ''
      model_kwargs: ''
      data:
        dataset: /mnt/ssd-2/lucia/datasets_local/train_64k.hf
        split: train
        subset: null
        prompt_column: text
        completion_column: ''
        conversation_column: ''
        reward_column: ''
        skip_nan_rewards: false
        truncation: false
        format_template: ''
        data_kwargs: ''
        chunk_length: 0
      tokenizer: ''
      drop_columns: true
      max_tokens: null
      use_tf32_matmuls: false
      debug: false
      projection_dim: 0
      include_bias: false
      reshape_to_square: false
      projection_type: rademacher
      projection_scale: jl
      projection_target: per_module
      token_batch_size: 1024
      max_batch_size: 64
      auto_batch_size: false
      optimizer_state: ''
      loss_fn: ce
      loss_reduction: sum
      label_smoothing: 0.0
      split_attention_modules: []
      attention:
        num_heads: 0
        head_size: 0
        head_dim: 0
      profile: false
      filter_modules: lm_head
      force_math_sdp: false
      attribute_tokens: false
      modules: []
    hessian_cfg:
      method: kfac
      ev_correction: true
      hessian_dtype: fp32
      use_dataset_labels: false
    score_cfg:
      query_path: ''
      score: individual
      batch_size: 1024
      query_batch_size: null
      precision: fp32
      modules: []
      higher_is_better: true
    preprocess_cfg:
      unit_normalize: false
      hessian_path: null
      ev_correction: false
      inversion_cfg:
        inversion: damped_inverse
        damping_factor: 0.1
      aggregation: none
      normalize_aggregated_grad: false
    hessian_pipeline_cfg:
      query:
        dataset: /mnt/ssd-2/lucia/datasets_local/query_20.hf
        split: train
        subset: null
        prompt_column: text
        completion_column: ''
        conversation_column: ''
        reward_column: ''
        skip_nan_rewards: false
        truncation: false
        format_template: ''
        data_kwargs: ''
        chunk_length: 0
      query_aggregation: none
      inversion_cfg:
        inversion: damped_inverse
        damping_factor: 0.1
      resume: false
metadata:
  bergson_version: null
  created: '2026-08-21T15:06:57+00:00'
  git_sha: 3c66bb51