Convex Markets / Datasets / Preference data / Intel Orca DPO Pairs
SKU PRF-3008 · Sold by External

Intel Orca DPO Pairs

Product specifications

SKUPRF-3008
Data typePreference data
Volume12,859 preference pairs (DPO triples: prompt + chosen + rejected)
Size on disk~36.3 MB (orca_rlhf.jsonl source file = 36,310,081 bytes; auto-converted Parquet ~19.1 MB)
FormatJSONL (single file orca_rlhf.jsonl; also auto-converted to Parquet on Hugging Face)
Access modelPUBLIC LICENSE
PricingFree · open-source license (Apache-2.0)
Quality score
LicenseApache License 2.0
Intel/orca_dpo_pairs is a preference-tuning dataset released by Intel that reformats instruction-response data from the Orca-style Open-Orca/OpenOrca dataset into paired preference records for Direct Preference Optimization. Each of the 12,859 rows contains four text fields: a system prompt, a question (the instruction/task), a chosen (preferred) response, and a rejected (dispreferred) response. The dataset is distributed as a single JSONL file (orca_rlhf.jsonl) under the Apache-2.0 license and is publicly downloadable without gating. The dataset card is minimal: it states only that the data 'contains 12k examples from Orca style dataset Open-Orca/OpenOrca' and references the Orca paper (arXiv:2306.02707). Note that the exact automated procedure used to assign chosen vs rejected (e.g. which model produced each side, or the judge used) is not documented on the primary dataset card.