theshark4/TRAILBench

Dataset Usage and Format

TRAILBench is released as JSON files, where each sample corresponds to one personalized function-calling evaluation instance. Each instance contains a current user query, the ground-truth function call, and the anonymized historical tool-use records available before the query time.

Sample-Level Fields

Each sample contains the following fields:

Field	Type	Description
sample_id	string	The unique identifier of the sample within the dataset file.
id	string	No specific function; mainly indicates the sample index for the current user
user_id	string	The identifie of which user the sample belongs to
time	string	The timestamp of the current user query. Only user-history records before this time are available to the model.
query	object	A set of natural language queries with different difficulty or ambiguity levels. All query variants correspond to the same ground-truth function call.
query["Minimum Difficulty"]	list of strings	The most explicit query form, where the key tool and parameter information is mostly stated by the user.
query["Low Difficulty"]	list of strings	A slightly less explicit query form with limited omitted information.
query["Medium Difficulty"]	list of strings	A more ambiguous query form that requires inferring some missing information from the user history.
query["High Difficulty"]	list of strings	The most ambiguous query form, where the model must infer key tool or parameter information from historical context.
answer	object	The ground-truth function call expected from the model.
answer["apitype"]	string	The domain or scenario of the target function call, such as transport, food, or reminder.
answer["toolname"]	string	The target tool or application to be selected, such as GaodeMap.
answer["apiname"]	string	The target API or function name, such as get_path.
answer["parameters"]	object	The ground-truth parameter values of the target function call.
answer["temporal_parameters"]	list of strings	The parameter names of temporal context personalized parameters. Omitted if no such parameters exist.
user_history	list or string	The anonymized chronological interaction history of the same user before the current query. In some raw files, this field may be stored as a JSON-encoded string and should be parsed with json.loads.

User-History Fields

The user_history field records previous tool-use behaviors of the same anonymized user. Each item in user_history corresponds to one historical function call.

Field	Type	Description
time	string	The timestamp of the historical interaction.
apitype	string	The domain or scenario of the historical tool call, such as transport, food, or entertainment.
toolname	string	The application or tool used in this historical interaction.
apiname	string	The API or function invoked in this historical interaction.
fulltoolname	string	The fully qualified function name in the format apitype.toolname.apiname. This field is provided for convenience and may duplicate information already contained in apitype, toolname, and apiname.
parameters	object	The parameters used in the historical function call.

Notes

The user_history field may contain interactions from multiple domains, including records that are not directly relevant to the current query. This design reflects realistic long-context user histories, where useful personalized signals are mixed with noisy behaviors.

For better compatibility with standard JSON parsers, we recommend converting user_history from a JSON-encoded string into a native JSON array during preprocessing.

More details

TRAILBench is the implementation for TRAILBench: Personalized Function Calling Benchmark Grounded in Real-world User Histories. It includes a high-quality dataset of anonymized interaction histories and an evaluation suite with controlled query vagueness.

supplementary_material_for_TRAILBench.pdf in our github repository include more dataset statistics and analysis in our paper like Dataset Details, Implemention details, Experiment analysis ,Prompt for generating and so on.

License

This entire project, including both the Codebase (all evaluation scripts, baseline implementations, and data processing pipelines) and the Dataset (the TRAILBench dataset, including anonymized user interaction logs and generated queries), is licensed under the Apache License 2.0.