Qwen3.6-27B Self-Data-Distillation: Single-Turn Reasoning Trajectories

Synthetic reasoning trajectories generated by the Qwen3.6-27B model via vLLM. The dataset contains four JSONL files, each line including a system prompt, user task, and the model's full output with embedded reasoning steps. It was created by author sleepyeldrazi and last updated on May 28, 2026.

Use Cases

• Training or fine-tuning language models based on the structured reasoning trajectories.
• Analyzing the reasoning patterns of a 27B-parameter model based on the embedded step-by-step outputs.
• Creating synthetic datasets for instruction-following tasks based on the prompt-response pairs.
• Studying self-data distillation methodologies based on the model's own generated trajectories.

Strengths

• Data includes the model's full reasoning steps embedded within the assistant content field.
• Trajectories are structured into four distinct categories, each in a separate JSONL file.
• Each record contains a complete system prompt, user task, and model output triplet.

Limitations

• Row count, file size, and column-level documentation are unknown, limiting suitability assessment.
• License information is unspecified, which may restrict usage.
• Data is synthetic and may reflect biases inherent to the generating model and its training data.

Provenance

Source

huggingface

Collection Method

Synthetically generated by running the Qwen3.6-27B model via vLLM.

Time Range

null

Freshness

Last updated 2026-05-28 13:48:51; freshness should be verified.

Geography

null