by OpenBMB
UltraData-SFT-Agent-2609 is OpenBMB's dataset of 483,661 agent trajectories for instruction tuning, covering tool use, search, code and general agent tasks. It keeps full trajectories with tool calls, observations and errors.
DatasetFree
Step 1: Open the dataset page on Hugging Face and pick the subset you need, such as Code-Agent.
Step 2: Load it with the datasets library using the openbmb/UltraData-SFT-Agent-2609 identifier and the subset name.
Step 3: Read the messages and tools fields of a few records to see the trajectory structure.
Step 4: Render the tool definitions into your target chat template before training.
Step 5: Apply the turn-level mask so excluded turns do not contribute to the SFT loss.
Each JSONL line is one trajectory with a uuid, a messages array of system, user, assistant and tool turns, the tool definitions, a source and a domain.
Some timeout, tool-error and recovery traces are kept after de-identification as auxiliary supervision, not as success demonstrations.
No. Some general and code tasks were sampled under multiple harnesses, so the totals count trajectories.
Trajectories go through format checks, tool-behavior checks, execution verification, deduplication, contamination screening, privacy filtering and multi-stage review including human spot checks.