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 oftoken_short
.p_long
: The model's prediction for the probability oftoken_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 oftoken_short
.long_max_id
: The id oftoken_long
.context
: The surrounding text of the prompt.context_ids
: The ids ofcontext
.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.