Spider DPO 1040: Preference Pairs for Text-to-SQL Model Training

Spider DPO 1040 is a compact training dataset containing 1,040 preference pairs for Direct Preference Optimization, derived from frontier-model disagreements on the Spider V1 benchmark. It also includes 7,000 supervised training examples from Spider formatted for use with LLaMA-Factory. The dataset was created by jk200201 and was last updated on July 5, 2026.

Use Cases

• Direct Preference Optimization (DPO) training for Text-to-SQL models based on the 1,040 preference pairs.
• Supervised fine-tuning of language models for SQL generation using the 7,000 formatted Spider train examples.
• Evaluating model alignment and preference learning techniques in the domain of structured query language generation.

Strengths

• Contains 1,040 preference pairs specifically curated for Direct Preference Optimization.
• Includes 7,000 supervised training examples from the established Spider benchmark.
• Dataset is associated with a specific, named companion model adapter (jk200201/qwen2.5-coder-7b-sql-dpo).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the preference pairs is explicitly stated, but the total dataset size and file formats are unknown.
• Description metadata is limited; the full description is hosted externally, and actual data quality requires manual inspection.

Provenance

Source

Derived from the Spider V1 benchmark.

Collection Method

Preference pairs were created from frontier-model disagreements on Spider V1.

Freshness

Last updated 2026-07-05 16:34:52; freshness should be verified.