ryoungj's picture
Update README.md
1ee629a verified
metadata
license: apache-2.0
task_categories:
  - text-generation
tags:
  - reasoning
  - latent-thoughts
  - latent-variable-models
  - expectation-maximization
  - self-improvement
  - mathematics
dataset_info:
  - config_name: finemath_4plus_final_2b_separated_bootstrap_split
    features:
      - name: text
        dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 8304174131
        num_examples: 1338556
    download_size: 3826592630
    dataset_size: 8304174131
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_1_mc_4
    features:
      - name: text
        dtype: string
      - name: num_chunks
        dtype: int64
      - name: chunks
        list:
          - name: suffix
            dtype: string
          - name: latent
            dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 35768747761
        num_examples: 1338556
    download_size: 17355271834
    dataset_size: 35768747761
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_2_mc_4
    features:
      - name: text
        dtype: string
      - name: num_chunks
        dtype: int64
      - name: chunks
        list:
          - name: suffix
            dtype: string
          - name: latent
            dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 35359265255
        num_examples: 1338556
    download_size: 17150348579
    dataset_size: 35359265255
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_3_mc_4
    features:
      - name: text
        dtype: string
      - name: num_chunks
        dtype: int64
      - name: chunks
        list:
          - name: suffix
            dtype: string
          - name: latent
            dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 35003567576
        num_examples: 1338556
    download_size: 16978801533
    dataset_size: 35003567576
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_4_mc_4
    features:
      - name: text
        dtype: string
      - name: num_chunks
        dtype: int64
      - name: chunks
        list:
          - name: suffix
            dtype: string
          - name: latent
            dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 34659154967
        num_examples: 1338556
    download_size: 16822737328
    dataset_size: 34659154967
  - config_name: finemath_4plus_final_500m_warmstart_split
    features:
      - name: text
        dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 2066140560
        num_examples: 334640
    download_size: 951823008
    dataset_size: 2066140560
  - config_name: finemath_4plus_final_500m_warmstart_split_with_gpt_latent_thoughts_latents
    features:
      - name: text
        dtype: string
      - name: num_chunks
        dtype: int64
      - name: chunks
        list:
          - name: suffix
            dtype: string
          - name: latent
            dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 8857963884
        num_examples: 334640
    download_size: 4239904449
    dataset_size: 8857963884
  - config_name: finemath_4plus_final_500m_warmstart_split_with_wrap_baseline_latents
    features:
      - name: text
        dtype: string
      - name: latent
        dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 14823881980
        num_examples: 1673200
    download_size: 7139488984
    dataset_size: 14823881980
  - config_name: finemath_4plus_final_500m_warmstart_split_with_wrap_cot_latents
    features:
      - name: text
        dtype: string
      - name: latent
        dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 13353860010
        num_examples: 1338560
    download_size: 6514552606
    dataset_size: 13353860010
  - config_name: finemath_4plus_final_fixed_val_split
    features:
      - name: text
        dtype: string
      - name: metadata
        dtype: string
    splits:
      - name: train
        num_bytes: 40464088
        num_examples: 6700
    download_size: 18501400
    dataset_size: 40464088
  - config_name: finetune_gsm8k
    features:
      - name: doc
        struct:
          - name: question
            dtype: string
          - name: answer
            dtype: string
      - name: text
        dtype: string
    splits:
      - name: test
        num_bytes: 2277260
        num_examples: 1319
      - name: train
        num_bytes: 11421745
        num_examples: 6725
      - name: validation
        num_bytes: 1275902
        num_examples: 748
    download_size: 7019889
    dataset_size: 14974907
  - config_name: finetune_math
    features:
      - name: doc
        struct:
          - name: problem
            dtype: string
          - name: solution
            dtype: string
          - name: answer
            dtype: string
          - name: level
            dtype: string
          - name: type
            dtype: string
      - name: text
        dtype: string
    splits:
      - name: test
        num_bytes: 12100367
        num_examples: 5000
      - name: train
        num_bytes: 17082822
        num_examples: 6750
      - name: validation
        num_bytes: 1847274
        num_examples: 750
    download_size: 14502759
    dataset_size: 31030463
configs:
  - config_name: finemath_4plus_final_2b_separated_bootstrap_split
    data_files:
      - split: train
        path: finemath_4plus_final_2b_separated_bootstrap_split/train-*
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_1_mc_4
    data_files:
      - split: train
        path: >-
          finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_1_mc_4/train-*
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_2_mc_4
    data_files:
      - split: train
        path: >-
          finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_2_mc_4/train-*
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_3_mc_4
    data_files:
      - split: train
        path: >-
          finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_3_mc_4/train-*
  - config_name: >-
      finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_4_mc_4
    data_files:
      - split: train
        path: >-
          finemath_4plus_final_2b_separated_bootstrap_split_with_bootstrap_latents_setup_bootstrap_fixed_data_iter_4_mc_4/train-*
  - config_name: finemath_4plus_final_500m_warmstart_split
    data_files:
      - split: train
        path: finemath_4plus_final_500m_warmstart_split/train-*
  - config_name: finemath_4plus_final_500m_warmstart_split_with_gpt_latent_thoughts_latents
    data_files:
      - split: train
        path: >-
          finemath_4plus_final_500m_warmstart_split_with_gpt_latent_thoughts_latents/train-*
  - config_name: finemath_4plus_final_500m_warmstart_split_with_wrap_baseline_latents
    data_files:
      - split: train
        path: >-
          finemath_4plus_final_500m_warmstart_split_with_wrap_baseline_latents/train-*
  - config_name: finemath_4plus_final_500m_warmstart_split_with_wrap_cot_latents
    data_files:
      - split: train
        path: >-
          finemath_4plus_final_500m_warmstart_split_with_wrap_cot_latents/train-*
  - config_name: finemath_4plus_final_fixed_val_split
    data_files:
      - split: train
        path: finemath_4plus_final_fixed_val_split/train-*
  - config_name: finetune_gsm8k
    data_files:
      - split: test
        path: finetune_gsm8k/test-*
      - split: train
        path: finetune_gsm8k/train-*
      - split: validation
        path: finetune_gsm8k/validation-*
  - config_name: finetune_math
    data_files:
      - split: test
        path: finetune_math/test-*
      - split: train
        path: finetune_math/train-*
      - split: validation
        path: finetune_math/validation-*

This dataset is associated with the paper Reasoning to Learn from Latent Thoughts. It contains data used for pretraining language models with a focus on improving data efficiency by modeling and inferring latent thoughts underlying the text generation process, such as on reasoning-intensive math corpus. An expectation-maximization algorithm is developed for models to self-improve their self-generated thoughts and data efficiency.