README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: token_short
      dtype: string
    - name: token_long
      dtype: string
    - name: p_short
      dtype: float32
    - name: p_long
      dtype: float32
    - name: JS
      dtype: float32
    - name: long_ids
      sequence: int32
    - name: short_max_id
      dtype: int64
    - name: long_max_id
      dtype: int64
    - name: context
      dtype: string
    - name: context_ids
      sequence: int32
    - name: p_delta_max
      dtype: float32
    - name: logit_excite_max
      dtype: float32
    - name: logit_inhibit_max
      dtype: float32
    - name: batch
      dtype: int64
    - name: sample
      dtype: int64
    - name: start
      dtype: int64
  splits:
    - name: scan
      num_bytes: 466393218
      num_examples: 1874497
  download_size: 0
  dataset_size: 466393218

scan_4

See https://confirmlabs.org/posts/catalog.html for details.

• text: The long prompt text with the the first token surrounded by square brackets.
• token_short: The model's prediction for the most likely token given the short prompt.
• token_long: The model's prediction for the most likely token given the long prompt.
• p_short: The model's prediction for the probability of token_short.
• p_long: The model's prediction for the probability of token_long.
• JS: The Jensen-Shannon divergence between the model's distribution over tokens given the short prompt and the model's distribution over tokens given the long prompt.
• long_ids: The ids of the tokens in the long prompt.
• short_max_id: The id of token_short.
• long_max_id: The id of token_long.
• context: The surrounding text of the prompt.
• context_ids: The ids of context.
• p_delta_max: The largest difference in probability for any token between the short and long prompt.
• logit_excite_max: The largest increase in logit for any token between the short and long prompt.
• logit_inhibit_max: The largest decrease in logit for any token between the short and long prompt.
• batch: The batch number of the prompt.
• sample: The sample number of the prompt.
• start: The start index of the prompt in the sample.