Qwen3.6-Plus High Reasoning Data for Model Distillation

1,697,229 tokens of instruction data were prepared for knowledge distillation using the Qwen3.6-Plus model. The dataset covers specialized topics including coding, mathematics, finance, medicine, and economics, with a maximum sequence length of 6,500 tokens per row. It was created by author 'khazarai' and was last updated in April 2026.

Use Cases

• Distill reasoning capabilities from the Qwen3.6-Plus teacher model into a smaller student model using the provided instruction-response pairs.
• Fine-tune a model for multi-domain question answering by leveraging the dataset's coverage of coding, mathematics, and finance topics.
• Create a specialized medical or economics chatbot by filtering and training on the relevant subsets of the instruction data.
• Benchmark model performance on long-context reasoning tasks using examples with sequences up to 6,500 tokens in length.

Strengths

• Contains 1,697,229 tokens of curated instruction data for distillation.
• Covers five distinct, complex domains: coding, mathematics, finance, medicine, and economics.
• Specifies a maximum sequence length of 6,500 tokens per data row, indicating support for long-context tasks.

Limitations

• The exact number of rows, column structure, and sample data are unknown, complicating assessment of data density and format.
• Potential for topic imbalance across the five listed domains without specific distribution metrics.
• Recency of the underlying source data and the teacher model's knowledge cutoff are unspecified.

Provenance

Source

huggingface, author khazarai

Collection Method

Prepared for distillation using outputs from the Qwen3.6-Plus teacher model.

Freshness

Last updated on 2026-04-05.