SlimOrca Dedup: 363k Deduplicated Instruction-Response Pairs

Open-Orca's SlimOrca Dedup is a dataset of 363,000 unique instruction-response examples derived from the SlimOrca collection. It was created by removing RLHF instances and applying minhash and Jaccard similarity techniques for deduplication. The dataset was last updated on Hugging Face on May 19, 2025.

Use Cases

• Training instruction-tuned language models based on the 363k unique examples.
• Benchmarking deduplication algorithms on synthetic instruction-response data.
• Studying the characteristics of language model outputs after RLHF instance removal.

Strengths

• Contains 363,000 unique examples, providing a substantial corpus.
• Explicitly excludes RLHF instances, which may reduce certain biases.
• Uses minhash and Jaccard similarity for systematic deduplication.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown beyond the stated 363k examples, which may limit suitability assessment.
• The description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Open-Orca via Hugging Face.

Collection Method

Derived from the SlimOrca dataset via deduplication and filtering.

Freshness

Last updated 2025-05-19 00:26:36.