|
--- |
|
annotations_creators: [] |
|
language: |
|
- en |
|
language_creators: [] |
|
license: [] |
|
multilinguality: |
|
- monolingual |
|
pretty_name: recipe RL roberta base |
|
size_categories: [] |
|
source_datasets: [] |
|
tags: [] |
|
task_categories: [] |
|
task_ids: [] |
|
--- |
|
|
|
# Dataset Description |
|
|
|
## Structure |
|
|
|
- Consists of 5 fields |
|
- Each row corresponds to a policy - sequence of actions, given an initial `<START>` state, and corresponding rewards at each step. |
|
|
|
## Fields |
|
|
|
`steps`, `step_attn_masks`, `rewards`, `actions`, `dones` |
|
|
|
## Field descriptions |
|
|
|
- `steps` (List of lists of `Int`s) - tokenized step tokens of all the steps in the policy sequence (here we use the `roberta-base` tokenizer, as `roberta-base` would be used to encode each step of a recipe) |
|
- `step_attn_masks` (List of lists of `Int`s) - Attention masks corresponding to `steps` |
|
- `rewards` (List of `Float`s) - Sequence of rewards (normalized b/w 0 and 1) assigned per step. |
|
- `actions` (List of lists of `Int`s) - Sequence of actions (one-hot encoded, as the action space is discrete). There are `33` different actions possible (we consider the maximum number of steps per recipe = `16`, so the action can vary from `-16` to `+16`; The class label is got by adding 16 to the actual action value) |
|
- `dones` (List of `Bool`) - Sequence of flags, conveying if the work is completed when that step is reached, or not. |
|
|
|
## Dataset Size |
|
|
|
- Number of rows = `2255673` |
|
- Maximum number of steps per row = `16` |
