Qodex Teacher Codex R2 Clean 48K: Code Trajectories for Teacher-Student Training

1146 cleaned ChatML rows of teacher-student code generation trajectories rendered with a 48k token budget. The dataset, created by bytkim, was last updated on May 7, 2026. It includes prompt/completion expansions for supervised fine-tuning tooling and detailed row-level provenance reports.

Use Cases

• Train supervised fine-tuning models based on the provided prompt/completion pairs.
• Analyze code generation trajectories based on the teacher-student interaction data.
• Benchmark model performance on code generation tasks using the cleaned ChatML format.
• Study data cleaning and provenance tracking methods for AI training datasets.

Strengths

• Includes 1146 cleaned ChatML rows of training data.
• Provides row-level source provenance and reports on dropped rows for transparency.
• Offers data in both ChatML and prompt/completion JSONL formats for different tooling needs.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary training split is known, but the total dataset scale and other splits are unknown.
• Description metadata is limited; actual data quality and content require manual inspection after download.

Provenance

Source

huggingface

Collection Method

Cleaned and merged from the Codex CLI teacher trajectory dataset.

Freshness

Last updated 2026-05-07 01:02:53; freshness should be verified.