CoDIT: Synthetic Responses for 250,333 Chat Instructions

CoDIT-LMSYS is a collection of 250,333 synthetic response texts generated for user instructions from the LMSYS-Chat-1M dataset. The data was synthesized using the CoDIT method by author Tatsuya-Ichinose, with specific versions generated by teacher models like Google's Gemma-3-27B-It and Qwen's Qwen3-8B. The dataset page was last updated on January 12, 2026.

Use Cases

• Fine-tuning language models for instruction-following based on the 250,333 synthesized instruction-response pairs.
• Training conversational AI agents using the synthetic chat data derived from LMSYS-Chat-1M.
• Benchmarking model performance on instruction-response tasks using data generated by specific teacher models like Gemma-3-27B-It.
• Studying the properties and effectiveness of the CoDIT synthesis method for generating training data.

Strengths

• Contains 250,333 instruction-response pairs, providing a substantial volume of training data.
• Offers multiple dataset versions synthesized by distinct, named teacher models (e.g., Gemma-3-27B-It, Qwen3-8B).
• Based on a known source dataset, LMSYS-Chat-1M, which provides context for the original user instructions.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual dataset splits (CoDIT-Gemma3, CoDIT-Qwen3-8B) is unknown, which may limit suitability assessment.

Provenance

Source

Synthesized from the LMSYS-Chat-1M dataset using the CoDIT method.

Collection Method

Response texts were synthesized using the CoDIT technique for the user instructions contained in the source dataset.

Freshness

Last updated 2026-01-12 09:51:04; freshness should be verified.