LongWorld
Collection
Long-context workflow probes, 128k baseline data, and selected SFT checkpoint exports. • 3 items • Updated
Datasets:
Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
License:
messages listlengths 3 3 | source stringclasses 3
values |
|---|---|
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
[{"content":"You are a careful analyst of long documents. Use only the provided context. If the cont(...TRUNCATED) | acc_search |
End of preview. Expand in Data Studio
LongWorld 128K Research Baselines
This research copy contains the acc, longtrace, and longmit related-work SFT datasets used with LongWorld. Each configuration has train and validation splits in Parquet format with OpenAI-style messages records.
The original row order, schemas, and message roles are preserved. Embedded authentication values and machine-specific operational paths or endpoints are replaced with redaction markers or portable placeholders. Public references, author attribution, and documented test and example content are retained; substitutions are recorded in the file manifest.
The original ACC and LongTrace exports were used in 680-step Base SFT runs. LongMIT is recipe-aligned and was not trained on 4B-Base in that round. These exports are separate from the LongWorld workflow dataset.
Load
from datasets import load_dataset
data = load_dataset("Xnhyacinth/LongWorld-128K-Baselines", "acc", split="train")
Sources and related resources
- ACC source dataset
- LongTraceRL source dataset
- LongMIT source dataset
- LongWorld workflow data
- LongWorld checkpoints
The file manifest and source identities are recorded in ARTIFACTS.json.
- Downloads last month
- 73