sldbench / README.md
pkuHaowei's picture
Add easy question scaling law dataset from ExpEmergence paper (U-shaped and Inverted-U Scaling behind Emergent Abilities of LLMs)
721b846 verified
metadata
dataset_info:
  - config_name: data_constrained_scaling_law
    features:
      - name: group
        dtype: string
      - name: unique_tokens
        dtype: float64
      - name: params
        dtype: float64
      - name: tokens
        dtype: float64
      - name: loss
        dtype: float64
    splits:
      - name: train
        num_bytes: 7084
        num_examples: 161
      - name: test
        num_bytes: 924
        num_examples: 21
    download_size: 7325
    dataset_size: 8008
  - config_name: domain_mixture_scaling_law
    features:
      - name: group
        dtype: string
      - name: proportion_domain_1
        dtype: float64
      - name: proportion_domain_2
        dtype: float64
      - name: proportion_domain_3
        dtype: float64
      - name: proportion_domain_4
        dtype: float64
      - name: proportion_domain_5
        dtype: float64
      - name: loss_domain_1
        dtype: float64
      - name: loss_domain_2
        dtype: float64
      - name: loss_domain_3
        dtype: float64
      - name: loss_domain_4
        dtype: float64
      - name: loss_domain_5
        dtype: float64
    splits:
      - name: train
        num_bytes: 7020
        num_examples: 80
      - name: test
        num_bytes: 2106
        num_examples: 24
    download_size: 15342
    dataset_size: 9126
  - config_name: easy_question_scaling_law
    features:
      - name: group
        dtype: string
      - name: log_flops
        dtype: float64
      - name: flops
        dtype: float64
      - name: brier_score
        dtype: float64
      - name: model_name
        dtype: string
    splits:
      - name: train
        num_bytes: 27304
        num_examples: 389
      - name: test
        num_bytes: 8530
        num_examples: 127
    download_size: 13707
    dataset_size: 35834
  - config_name: lr_bsz_scaling_law
    features:
      - name: group
        dtype: string
      - name: lr
        dtype: float64
      - name: bsz
        dtype: float64
      - name: data_size
        dtype: float64
      - name: non_embedding_param_size
        dtype: float64
      - name: lm_loss
        dtype: float64
    splits:
      - name: train
        num_bytes: 140504
        num_examples: 2702
      - name: test
        num_bytes: 6084
        num_examples: 117
    download_size: 39110
    dataset_size: 146588
  - config_name: lr_bsz_scaling_law_modified
    features:
      - name: group
        dtype: string
      - name: lr
        dtype: float64
      - name: bsz
        dtype: float64
      - name: data_size
        dtype: float64
      - name: non_embedding_param_size
        dtype: float64
      - name: lm_loss
        dtype: float64
    splits:
      - name: train
        num_bytes: 186836
        num_examples: 3593
      - name: test
        num_bytes: 6084
        num_examples: 117
    download_size: 44596
    dataset_size: 192920
  - config_name: moe_scaling_law
    features:
      - name: group
        dtype: string
      - name: num_experts
        dtype: float64
      - name: dense_parameter_count
        dtype: float64
      - name: loss_validation
        dtype: float64
    splits:
      - name: train
        num_bytes: 6948
        num_examples: 193
      - name: test
        num_bytes: 1008
        num_examples: 28
    download_size: 6463
    dataset_size: 7956
  - config_name: parallel_scaling_law
    features:
      - name: num_params
        dtype: int64
      - name: parallel_size
        dtype: int64
      - name: group
        dtype: string
      - name: loss
        dtype: float64
    splits:
      - name: train
        num_bytes: 1170
        num_examples: 36
      - name: test
        num_bytes: 390
        num_examples: 12
    download_size: 4551
    dataset_size: 1560
  - config_name: sft_scaling_law
    features:
      - name: group
        dtype: string
      - name: sft_data_size
        dtype: int64
      - name: sft_loss
        dtype: float64
    splits:
      - name: train
        num_bytes: 25896
        num_examples: 504
      - name: test
        num_bytes: 2158
        num_examples: 42
    download_size: 9274
    dataset_size: 28054
  - config_name: vocab_scaling_law
    features:
      - name: group
        dtype: string
      - name: non_vocab_parameters
        dtype: float64
      - name: vocab_size
        dtype: float64
      - name: num_characters
        dtype: float64
      - name: unigram_normalized_loss
        dtype: float64
    splits:
      - name: train
        num_bytes: 47520
        num_examples: 1080
      - name: test
        num_bytes: 5280
        num_examples: 120
    download_size: 23841
    dataset_size: 52800
configs:
  - config_name: data_constrained_scaling_law
    data_files:
      - split: train
        path: data_constrained_scaling_law/train-*
      - split: test
        path: data_constrained_scaling_law/test-*
  - config_name: domain_mixture_scaling_law
    data_files:
      - split: train
        path: domain_mixture_scaling_law/train-*
      - split: test
        path: domain_mixture_scaling_law/test-*
  - config_name: easy_question_scaling_law
    data_files:
      - split: train
        path: easy_question_scaling_law/train-*
      - split: test
        path: easy_question_scaling_law/test-*
  - config_name: lr_bsz_scaling_law
    data_files:
      - split: train
        path: lr_bsz_scaling_law/train-*
      - split: test
        path: lr_bsz_scaling_law/test-*
  - config_name: lr_bsz_scaling_law_modified
    data_files:
      - split: train
        path: lr_bsz_scaling_law_modified/train-*
      - split: test
        path: lr_bsz_scaling_law_modified/test-*
  - config_name: moe_scaling_law
    data_files:
      - split: train
        path: moe_scaling_law/train-*
      - split: test
        path: moe_scaling_law/test-*
  - config_name: parallel_scaling_law
    data_files:
      - split: train
        path: parallel_scaling_law/train-*
      - split: test
        path: parallel_scaling_law/test-*
  - config_name: sft_scaling_law
    data_files:
      - split: train
        path: sft_scaling_law/train-*
      - split: test
        path: sft_scaling_law/test-*
  - config_name: vocab_scaling_law
    data_files:
      - split: train
        path: vocab_scaling_law/train-*
      - split: test
        path: vocab_scaling_law/test-*