WIA-Corpus-RL

File Purpose and Sample Count

Each line is one sample containing input context and a target output.

File	Purpose	Samples	Top-level fields
train.jsonl	World-model reinforcement-learning training data	134,555	6 base fields; 16,496 samples also have 6 additional fields, for 12 total
train-filtered.jsonl	Filtered training data	108,761	6 base fields; 15,709 samples also have 6 additional fields, for 12 total
val.jsonl	Validation data	726	6 base fields

The Viewer default configuration uses train.jsonl; the filtered configuration uses train-filtered.jsonl. Both configurations use val.jsonl for the validation split.

Field Definitions

The six base fields are present in every sample in every file:

Field	Type	Purpose
system	string	System instruction defining the model role and prediction rules.
prompt	string	Input context, including tasks, previous actions, or environment observations.
ground_truth	string	Target output or next environment observation.
source	string	Sample source identifier.
domain	string	Task domain.
dataset	string	Source dataset name.

The following six additional fields appear only in some training samples:

Field	Type	Purpose
source_record_index	integer	Index of the original record.
source_turn_index	integer	Turn index in the original record.
system_tokens	integer	Token count of system.
prompt_tokens	integer	Token count of prompt.
ground_truth_tokens	integer	Token count of ground_truth.
total_tokens	integer	Total token count for the sample.

Download

pip install -U huggingface_hub

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="LightningCreeper/WIA-Corpus-RL",
    repo_type="dataset",
    allow_patterns=["train.jsonl", "train-filtered.jsonl", "val.jsonl"],
    local_dir="WIA-Corpus-RL",
)