Dataset

synthlabs/rl-preferences

1.2k

Pairwise preference labels generated and verified by the Synth optimizer.

rlhfpreferencessynthetic

480k
rows: 6.7 GB
size: 22.4k
downloads: 44.0k
eval runs: 9h ago
updated

Dataset card

Pairwise preference labels generated and verified by the Synth optimizer.

Records are deduplicated and license-checked before publishing. Each row carries its source trajectory id, so any example can be traced back to the run, task, and verifier score that produced it.

Format: JSONL / Parquet
Schema: trajectory-v2
Splits: train · eval
Provenance: Verified trials

Use this dataset

CLI

$tune pull synthlabs/rl-preferences

Python

from tune import datasets

ds = datasets.load("synthlabs/rl-preferences", split="train")
for row in ds.take(3):
    print(row["trajectory_id"], row["reward"])