DeliveryBench / README.md
fstxwd's picture
Release version-matched evaluation labels and runnable scoring resources
a471c50 verified
|
Raw History Blame Contribute Delete
8.91 kB
metadata
pretty_name: DeliveryBench
language:
  - zh
size_categories:
  - 100K<n<1M
tags:
  - agents
  - personalization
  - recommendation
  - food-ordering
  - longitudinal
configs:
  - config_name: catalogs
    default: true
    data_files:
      - split: full
        path: catalog_*.parquet
  - config_name: users
    data_files:
      - split: full
        path: users.parquet
  - config_name: requests
    data_files:
      - split: full
        path: requests.parquet
  - config_name: locations
    data_files:
      - split: full
        path: locations.parquet
  - config_name: evaluation_labels
    data_files:
      - split: train
        path: evaluation_labels_train.parquet
      - split: validation
        path: evaluation_labels_validation.parquet
      - split: test
        path: evaluation_labels_test.parquet
  - config_name: training_history
    data_files:
      - split: train
        path: evaluation_training_history.parquet
  - config_name: request_context
    data_files:
      - split: full
        path: evaluation_request_context.parquet

DeliveryBench

DeliveryBench supports personalized food-ordering agents through real, location-specific menus, constructed user profiles and longitudinal natural-language requests. It covers Guangzhou, Shenzhen, Shanghai and Beijing.

Available data: 1,000 synthetic user profiles, 8,568 eligible ordering requests and 722,975 menu configurations with product attributes and selected specification options, covering 20 episode-linked delivery locations.

Run evaluation

The evaluation guide provides download, install, train-prior, agent-run and scoring commands. Download the accompanying ada_food_benchmark-0.2.0-py3-none-any.whl and evaluation_tools-0.2.0.zip for the framework and scorer. The source archive includes tests and a custom-agent interface.

The evaluation_labels configuration has 5,938 / 1,239 / 1,391 train/validation/test targets; training_history contains the 5,938 actual training choices. request_context preserves relative chronology for all requests. IDs and the existing catalog/input table contents are unchanged.

from datasets import load_dataset

targets = load_dataset("fstxwd/DeliveryBench", "evaluation_labels")
history = load_dataset("fstxwd/DeliveryBench", "training_history", split="train")

The current runnable track uses ada-bench-strict-config-v1. Historical paper experiments used a different simulator/judge; paper_metric_reanalysis.zip separately provides sanitized recorded metrics and point-estimate reanalysis. Its 144 backend episodes use synthetic profiles, not human participants. Public test targets make this an open-label reproducibility release.

Dataset scale

City Users Eligible requests Menu configurations Catalog locations
Guangzhou 250 2,305 139,724 5
Shenzhen 250 1,916 223,439 5
Shanghai 250 2,110 232,427 5
Beijing 250 2,237 127,385 5
Total 1,000 8,568 722,975 20

The original constructed corpus contains 8,873 episodes; 8,568 passed the menu-feasibility audit and appear in this release. Each city contributes 250 user profiles. The split column retains the original user-disjoint assignment.

Partition Users Eligible requests
Training 700 5,938
Validation 150 1,239
Test 150 1,391

The broader collection targets 200 locations, with 198 completed, and contains 64,095 complete menu snapshots, 4,926,925 product observations and 35,052,171 specification-option observations. Those broader collection counts are not the downloadable catalog's size. This release contains the audited 722,975-configuration catalog at the 20 episode-linked locations. Configurations and location-specific observations are not counts of unique dishes or unique businesses.

Files and joins

Files Contents Rows
users.parquet; users.jsonl User ID, city, partition and constructed food/drink preferences 1,000 each
requests.parquet; requests.jsonl Natural-language requests, sequence, time context, partition and catalog location 8,568 each
locations.parquet; locations.jsonl Anonymous location ID, city, catalog count, request count and collection status 20 each
catalog_guangzhou.parquet, catalog_shenzhen.parquet, catalog_shanghai.parquet, catalog_beijing.parquet Location-specific menu configurations, product information, attributes and selected options 722,975 combined
catalog_manifest.json Per-file checksums, row counts, attribute statistics, missingness and integrity checks —
preview_manifest.json Versioned release scope and file checksums —

Parquet and JSONL copies of the small input tables contain the same records. Requests join to profiles through user_id and to the catalog through location_id. Anonymous location IDs preserve these joins without publishing addresses or coordinates. sequence_index preserves request order within each user. Existing user/request IDs are retained from the input-data release.

Product configuration fields

Each catalog row represents one enumerated configuration from a location-specific menu snapshot.

Field Meaning
configuration_id Unique opaque configuration identifier
location_id, city Anonymous catalog location and city
shop_id, item_id, sku_id Opaque, location-scoped merchant, item and SKU identifiers
name, shop_name, brand Product name, merchant name and recorded brand
family, domain Processed food-family and food/drink classification
sku_name Source SKU label, such as size or portion
attributes Normalized menu-derived {name, value} pairs, including available size, sweetness, ice level or spiciness fields
options Selected source-menu {group, value, option_id} entries, preserving labels such as temperature, sweetness and add-ins
price_cents, price_scope Recorded base/SKU price in integer cents and its scope
rating, sales, eta, distance Available provider service-quality fields; missing evidence remains null
checkout_verified Whether the source configuration was verified at checkout

Menu labels and normalized attributes describe the available menu evidence. They are not independent ingredient, allergy or preparation guarantees. price_scope distinguishes base/SKU prices from checkout totals; selected extras and final prices may remain unverified. Provider quality fields retain their source values and missingness; no common cross-provider scale or unit conversion is asserted.

The catalog preserves the source enumeration bounds. bounded_complete in the location table denotes completion within recorded collection/enumeration bounds, not a guarantee that every platform merchant or possible option combination was obtained. Opaque IDs support snapshot-local joins, not global merchant identity resolution.

Users and requests

Profiles describe constructed preferences rather than identified people. Real ordering traces inform activity, timing and expenditure context; semantic food-family and attribute preferences are separately constructed and grounded in menu supply. Natural-language requests are synthetic ordering inputs, not real customer conversations.

Profile preferences contain domain/family entries and personal attribute name/value pairs. hour is 0–23; weekday retains the source 0–6 encoding, and period retains the source categorical code. task_type preserves the original request category.

Raw API responses, credentials, precise addresses/coordinates and source-user linkage hashes are excluded. Contact-like strings in menu text are removed. Version-matched targets for all 8,568 requests, 5,938 actual training choices and relative request timestamps are now included. Test answers are intentionally public for open-label reproducibility. Profile fields supplied for inspection are excluded from permitted observations by the evaluation loader.

Loading

from datasets import load_dataset

catalogs = load_dataset("fstxwd/DeliveryBench", "catalogs", split="full")
users = load_dataset("fstxwd/DeliveryBench", "users", split="full")
requests = load_dataset("fstxwd/DeliveryBench", "requests", split="full")
locations = load_dataset("fstxwd/DeliveryBench", "locations", split="full")

assert len(catalogs) == 722975
assert len(users) == 1000 and len(requests) == 8568

request = requests[0]
local_candidates = catalogs.filter(
    lambda row: row["location_id"] == request["location_id"]
)
print(request["message"])
print(local_candidates[0])

For a specific partition, filter the input tables by their split column. The catalog is partitioned by city for storage; location_id identifies the candidate pool for a request. No purchase submission or payment is needed to inspect these files.