Qwen 3.6 Plus: Multi-Turn Agent Tool-Calling Trajectories for SFT

Multi-turn tool-calling trajectories generated by the Qwen 3.6 Plus model via OpenRouter on the ToolScale platform. The dataset includes both passing and near-passing rollouts, with each row being a flattened conversation prefix ending at an assistant turn, ready for next-token supervised fine-tuning. Author ansulev uploaded this dataset to Hugging Face, with a last recorded update on 2026-05-15.

Use Cases

• Supervised fine-tuning of language models for tool-calling based on multi-turn trajectories.
• Analyzing model reasoning patterns based on the included reasoning_content field.
• Training or evaluating agentic models using a mix of passing and near-passing rollouts.
• Studying the quality thresholds of agent behavior using the provided reward and score signals.

Strengths

• Includes both passing and near-passing rollouts, allowing for flexible quality threshold selection.
• Assistant turns contain a reasoning_content field, providing insight into the model's internal process.
• Data is formatted as flattened conversation prefixes, making it ready for next-token prediction tasks.

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 is unknown, which may limit suitability assessment.

Provenance

Source

Generated by Qwen 3.6 Plus via OpenRouter on the ToolScale platform.

Collection Method

Multi-turn tool-calling trajectories generated by the model, likely for agent training.

Freshness

Last updated 2026-05-15 22:57:50; freshness should be verified.