Dataset

UltraData-SFT-Agent-2609

Contains ~484k text-only agent trajectories across 4 agent directions for supervised fine-tuning of tool-using and multi-step assistant behavior.

Hugging Face ↗Aug 31, 2026Paper ↗

  • llm
  • nlp
  • agent-trajectories
  • tool-use
  • function-calling
Six-stage pipeline diagram showing the data construction method
Six-stage pipeline diagram showing the data construction method

What it is

Agent SFT dataset with ~484k executable text trajectories across tool use, search, coding, and general workflows, including tool calls, observations, errors, retries, and final answers. ~484k text agent trajectories across 4 configs: General-Agent, Tool-Use, Code-Agent, and Search-Agent. Each JSONL sample stores full multi-turn messages, tool schemas, tool returns, environment feedback, errors, retries, and final outputs.

Key results

  • 483,661 trajectories in JSONL format
  • 4 configs: General, Tool-Use, Code, Search
  • Apache-2.0 license; English and Chinese

Useful for supervised fine-tuning LLM agents on tool calling, search, coding, and recovery from execution errors.

One useful ML email a week.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.