metadata
dataset_info:
features:
- name: is_thoughtful
dtype: bool
- name: uuid
dtype: string
- name: chunk_index
dtype: int64
- name: chunk_start_index
dtype: int64
- name: chunk_end_index
dtype: int64
- name: num_scored_tokens
dtype: int64
- name: input_ids
list:
list: int32
- name: thought_text
list: large_string
- name: g
list: int64
- name: truncated
list: bool
- name: thoughtless_log_prob_sum
dtype: float32
- name: thoughtless_log_prob_per_token
list: float32
- name: thoughtful_log_prob_sum
list: float32
- name: thoughtful_log_prob_per_token
list:
list: float32
splits:
- name: train
num_bytes: 93960828.5
num_examples: 311296
- name: test
num_bytes: 93960828.5
num_examples: 311296
download_size: 69936144
dataset_size: 187921657
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
Example of a chunk record dataset, made for debugging purposes.
Made by taking the step-1216 validation chunk records of reason-only-k=32-lr=0.00015-replay=0.5 and splitting the doc uuids in half into train and test. Thus, all of a document's chunks lie within exactly one of the splits.
There are (4096/32) * 4864 =622_592$ chunks (including the first chunk) collectively.
| Split | Chunks | Thoughtful Chunks | Thoughtful Proportion |
|---|---|---|---|
train |
311_296 | 38_557 | 0.1239 |
test |
311_296 | 38_659 | 0.1242 |
| Total | 622_592 | 77_216 | 0.1240 |