id stringlengths 10 46 | question listlengths 1 1 | involved_classes listlengths 1 6 | verifier_type stringlengths 13 32 ⌀ | initial_config stringlengths 464 109k | flakiness_config stringclasses 0
values |
|---|---|---|---|---|---|
ptc_grocery_gift_card_0 | [
[
{
"role": "user",
"content": "I have gift card balances at several grocery delivery stores. I want to buy everything on my shopping list while spending as little cash out-of-pocket as possible. Each store I order from charges a delivery fee on top of item prices. A gift card covers part or all of ... | [
"GroceryDeliveryAPI"
] | grocery_gift_card | {"GroceryDeliveryAPI": {"shopping_list": ["eggs", "milk", "bread", "chicken", "rice", "avocados", "pasta", "olive oil", "butter", "cheese", "yogurt", "salmon", "broccoli", "tomatoes", "onions", "garlic", "potatoes", "carrots", "apples", "bananas", "orange juice", "cereal", "steak", "wine", "salad kit", "ice cream", "al... | null |
ptc_grocery_set_cover_0 | [
[
{
"role": "user",
"content": "I need to get every item on my shopping list delivered, but no single store carries everything. Each store I order from charges a separate delivery fee. Find the smallest group of stores that together stock every item on my list. If there are ties in the number of sto... | [
"GroceryDeliveryAPI"
] | grocery_set_cover | {"GroceryDeliveryAPI": {"shopping_list": ["eggs", "milk", "bread", "chicken", "rice", "avocados", "pasta", "olive oil", "butter", "cheese", "yogurt", "salmon", "broccoli", "tomatoes", "onions", "garlic", "potatoes", "carrots", "apples", "bananas", "orange juice", "cereal", "steak", "wine", "salad kit", "ice cream", "al... | null |
ptc_grocery_single_vs_split_0 | [
[
{
"role": "user",
"content": "I want to buy everything on my shopping list. Compare two options: (A) ordering everything from whichever single store gives me the lowest total cost including item prices, delivery fee, and service fee; versus (B) splitting my order across exactly three stores to min... | [
"GroceryDeliveryAPI"
] | grocery_single_vs_split | {"GroceryDeliveryAPI": {"shopping_list": ["eggs", "milk", "bread", "chicken", "rice", "avocados", "pasta", "olive oil", "butter", "cheese", "yogurt", "salmon", "broccoli", "tomatoes", "onions", "garlic", "potatoes", "carrots", "apples", "bananas", "orange juice", "cereal", "steak", "wine", "salad kit", "ice cream", "al... | null |
ptc_multi_group_dinner_0 | [[{"role":"user","content":"My friends and I are choosing between six restaurants for dinner tonight(...TRUNCATED) | [
"YelpAPI",
"GoogleMapsAPI",
"UberAPI"
] | multi_group_dinner | "{\"YelpAPI\": {\"profile\": {\"name\": \"Alex Chen\", \"email\": \"alex.chen@email.com\", \"locatio(...TRUNCATED) | null |
ptc_multi_meeting_point_0 | [[{"role":"user","content":"My friend and I are coming from different neighborhoods and want to meet(...TRUNCATED) | [
"YelpAPI",
"GoogleMapsAPI",
"UberAPI"
] | multi_meeting_point | "{\"YelpAPI\": {\"profile\": {\"name\": \"Alex Chen\", \"email\": \"alex.chen@email.com\", \"locatio(...TRUNCATED) | null |
ptc_org_chart_largest_manager_0 | [[{"role":"user","content":"Look through the org chart history and tell me which manager has supervi(...TRUNCATED) | [
"OrgChartAPI"
] | exact_json_match | "{\"OrgChartAPI\": {\"employees\": [{\"employee_id\": \"e1\", \"employee_name\": \"Alice\"}, {\"empl(...TRUNCATED) | null |
ptc_org_chart_most_teammates_0 | [[{"role":"user","content":"Look through this project history and tell me which person has worked wi(...TRUNCATED) | [
"OrgChartAPI"
] | exact_json_match | "{\"OrgChartAPI\": {\"employees\": [{\"employee_id\": \"e1\", \"employee_name\": \"Alice\"}, {\"empl(...TRUNCATED) | null |
ptc_queries_calendar_compact_oneononones_0 | [[{"role":"user","content":"Schedule four 30-minute one-on-ones for me with Alice, Bob, Charlie, and(...TRUNCATED) | [
"EnterpriseCalendarAPI"
] | compact_scheduling | "{\"EnterpriseCalendarAPI\": {\"metadata\": {\"simulated_now_iso\": \"2025-05-18T20:00:00\", \"simul(...TRUNCATED) | null |
ptc_queries_calendar_five_attendees_0 | [[{"role":"user","content":"Next week, give a possible 45-minute meeting time for Alice, Bob, Charli(...TRUNCATED) | [
"EnterpriseCalendarAPI"
] | slot_feasibility | "{\"EnterpriseCalendarAPI\": {\"metadata\": {\"simulated_now_iso\": \"2025-03-19T14:00:00\", \"simul(...TRUNCATED) | null |
ptc_queries_calendar_fragmented_weekday_0 | [[{"role":"user","content":"Look at my calendar for the next four weeks and tell me which day of the(...TRUNCATED) | [
"EnterpriseCalendarAPI"
] | exact_json_match | "{\"EnterpriseCalendarAPI\": {\"metadata\": {\"simulated_now_iso\": \"2025-05-04T20:00:00\", \"simul(...TRUNCATED) | null |
End of preview. Expand in Data Studio
PTC-Bench
A benchmark for evaluating LLM agents' programmatic tool calling (PTC): agents solve multi-step tasks over simulated APIs (calendar, org chart, Yelp, Uber, Google Maps, grocery delivery, DoorDash, Venmo, Spotify), either through direct tool calls or by writing code that calls the APIs.
Splits
| Split | Tasks | Description |
|---|---|---|
main |
100 | Single-turn tasks |
error_recovery |
70 | Single-turn tasks with injected API failures (flakiness_config, or the *Alt API variants that fail on the 13th call) |
multi_turn |
70 | Two-turn tasks: a follow-up user turn refines the first request |
Fields
| Field | Type | Description |
|---|---|---|
id |
string | Task id |
question |
list of turns, each a list of {role, content} messages |
The user turn(s) |
involved_classes |
list of strings | Simulated API classes the task uses |
verifier_type |
string or null | Verifier label (metadata only; the evaluator dispatches verifiers by id; null on 30 main tasks) |
initial_config |
string (JSON) | Initial state of each simulated API, keyed by class name |
flakiness_config |
string (JSON) or null | Error-injection schedule (error_recovery only) |
initial_config and flakiness_config are stored as JSON strings because their structure
differs by API; json.loads returns the original object.
Usage
import json
from datasets import load_dataset
ds = load_dataset("shuo-yuan/PTC-Bench", split="main")
task = ds[0]
state = json.loads(task["initial_config"])
Scoring requires the simulated API implementations and per-task verifiers from the PTC-Bench code repository; the data here contains the tasks and initial API states only.
License
The dataset is released under the Apache License 2.0, the same license as the PTC-Bench code. All API states (reviews, calendars, org charts, transactions, playlists) are synthetic.
- Downloads last month
- 4