Datasets:

sleepyeldrazi
/

qwen3.6-27b-self-data-distillation-dataset

id stringlengths 23 27	source stringclasses 1 value	task stringlengths 49 1.15k	messages listlengths 3 3	input_tokens int64 32 599	output_tokens int64 1.74k 16.4k
traj_1779884057103_17_73106	synthetic-qwen3.6-27b	Write a pipeline prompt for a tool that must decide when to search, when to calculate, and when to respond directly based on user intent.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Write a pipeline prompt for a tool that must decide when to search, when to calculate, and when to respond directly based on user intent." }, { "role": "assistant", "content": "Here's a ...	50	2,957
traj_1779884058980_3_36051	synthetic-qwen3.6-27b	Design a customer support notification agent that monitors events, applies routing rules, and delivers alerts through appropriate channels.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Design a customer support notification agent that monitors events, applies routing rules, and delivers alerts through appropriate channels." }, { "role": "assistant", "content": "Here's ...	43	2,971
traj_1779884088400_4_67448	synthetic-qwen3.6-27b	Design an agent that follows the 'decompose-solve-compose' pattern: the agent first decomposes data, then solves it, and finally composes the result. Specify the tools and state management for each phase.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Design an agent that follows the 'decompose-solve-compose' pattern: the agent first decomposes data, then solves it, and finally composes the result. Specify the tools and state management for each ...	65	3,125
traj_1779884119962_9_68674	synthetic-qwen3.6-27b	Design a tool that extracts structured medical report data from unstructured text. Specify the output schema with all fields: patient_id, diagnosis, medications, test_results, recommendations, follow_up_date. Include validation rules.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Design a tool that extracts structured medical report data from unstructured text. Specify the output schema with all fields: patient_id, diagnosis, medications, test_results, recommendations, follo...	64	3,403
traj_1779884156740_13_33314	synthetic-qwen3.6-27b	Design an agent that follows the 'parse-transform-load' pattern: the agent first parses data, then transforms it, and finally loads the result. Specify the tools and state management for each phase.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Design an agent that follows the 'parse-transform-load' pattern: the agent first parses data, then transforms it, and finally loads the result. Specify the tools and state management for each phase....	61	3,480
traj_1779884186725_0_35961	synthetic-qwen3.6-27b	Build a test suite for an agent that uses fallback chain: try primary, then secondary, then cached. Cover success paths, partial failures, and full failures.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Build a test suite for an agent that uses fallback chain: try primary, then secondary, then cached. Cover success paths, partial failures, and full failures." }, { "role": "assistant", "...	54	3,941
traj_1779884189096_8_99661	synthetic-qwen3.6-27b	Build a timeout recovery for an LLM-based analysis tool. The call hangs for 30 seconds. How does the agent cancel, retry with fewer tokens, and proceed?	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Build a timeout recovery for an LLM-based analysis tool. The call hangs for 30 seconds. How does the agent cancel, retry with fewer tokens, and proceed?" }, { "role": "assistant", "conte...	57	3,951
traj_1779884197430_36_28074	synthetic-qwen3.6-27b	Design a personal finance agent that connects to bank APIs, categorizes transactions, detects unusual spending, generates budgets, and sends weekly reports. Describe the data flow.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Design a personal finance agent that connects to bank APIs, categorizes transactions, detects unusual spending, generates budgets, and sends weekly reports. Describe the data flow." }, { "ro...	54	3,174
traj_1779884199115_28_1009	synthetic-qwen3.6-27b	Build a multi-turn conversation agent that maintains state across tool calls: remembering previous search results, user preferences, and pending actions.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Build a multi-turn conversation agent that maintains state across tool calls: remembering previous search results, user preferences, and pending actions." }, { "role": "assistant", "cont...	47	3,373
traj_1779884199910_19_19331	synthetic-qwen3.6-27b	Build a travel workflow automation agent that sequences multi-step processes with approval gates at each stage.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Build a travel workflow automation agent that sequences multi-step processes with approval gates at each stage." }, { "role": "assistant", "content": "Here's a thinking process:\n\n1. *...	40	3,489
traj_1779884204082_39_19681	synthetic-qwen3.6-27b	Build a legal customer service agent that handles account inquiries, processes requests, and escalates complex issues to human agents. Define the tool set and decision logic.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Build a legal customer service agent that handles account inquiries, processes requests, and escalates complex issues to human agents. Define the tool set and decision logic." }, { "role": "...	53	3,161
traj_1779884410970_46_84879	synthetic-qwen3.6-27b	Write a prompt template for a tool that has access to a code interpreter, file system, and web browser.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Write a prompt template for a tool that has access to a code interpreter, file system, and web browser." }, { "role": "assistant", "content": "Here's a thinking process:\n\n1. **Analyze...	44	2,184
traj_1779884447981_43_84202	synthetic-qwen3.6-27b	Build a HR customer service agent that handles account inquiries, processes requests, and escalates complex issues to human agents. Define the tool set and decision logic.	[ { "role": "system", "content": "You are a helpful AI assistant." }, { "role": "user", "content": "Build a HR customer service agent that handles account inquiries, processes requests, and escalates complex issues to human agents. Define the tool set and decision logic." }, { "role": "ass...	53	2,581

End of preview. Expand in Data Studio

Qwen3.6-27B Self-Data-Distillation Trajectories

Single‑turn reasoning trajectories generated by running Qwen3.6‑27B (via vLLM). Each trajectory contains a system prompt, a user task, and the model's full output (including reasoning steps embedded in the assistant content field).

Data Format

Four JSONL files, one per category. Each line is:

{
  "id": "traj_<timestamp>_<idx>_<seq>",
  "source": "synthetic-qwen3.6-27b",
  "task": "<the prompt given to the model>",
  "messages": [
    {"role": "system",   "content": "You are a helpful AI assistant."},
    {"role": "user",     "content": "<task>"},
    {"role": "assistant","content": "<full response with reasoning>", "reasoning_content": ""}
  ],
  "input_tokens":  <int>,
  "output_tokens": <int>
}

Generation settings

Parameter	Value
Model	Qwen3.6‑27B
Temperature	0.7
Max tokens	16,384
Concurrency	256 per category
System prompt	`"You are a helpful AI assistant."`
Source	synthetic-qwen3.6-27b

Distribution

Category	Samples	% of Dataset	Total Tokens	% of Tokens	Avg Tokens	Avg In	Avg Out	Min Tokens	Max Tokens
Math	399	9.1%	4,245,768	21.4%	10,641	141	10,500	1,887	16,679
World Knowledge	2,000	45.5%	8,000,692	40.2%	4,000	82	3,918	2,396	5,492
Coding	600	13.6%	2,404,965	12.1%	4,008	37	3,971	1,802	5,614
Agentic	1,400	31.8%	5,229,495	26.3%	3,735	50	3,685	1,947	5,482
Total	4,399	100.0%	19,880,920	100.0%	4,520	69	4,451	1,802	16,679

Note: Math trajectories are longer on average (10.6K tokens) because the model produces detailed step‑by‑step reasoning. World Knowledge, Coding, and Agentic trajectories are shorter (~3.7–4K tokens). The max token ceiling of 16,384 was only approached in Math.

Files

File	Size
`data/math_trajectories.jsonl`	11 MB
`data/world_knowledge_trajectories.jsonl`	34 MB
`data/coding_trajectories.jsonl`	9.8 MB
`data/agentic_trajectories.jsonl`	22 MB

Prompt Sources

Category	Source
Math	nvidia/OpenMathReasoning
World Knowledge	soldni/jeopardy + Tier1 questions
Coding	Tier1 + generated coding tasks
Agentic	Tier1 + generated agentic/tool‑use scenarios

License

Apache 2.0

Downloads last month: 106