Datasets:

amergoel
/

SoftOR

language:
  - en
license: other
license_name: softor-mixed-source-terms
license_link: LICENSE.md
task_categories:
  - text-generation
size_categories:
  - n<1K
tags:
  - optimization
  - operations-research
  - mathematical-modeling
  - soft-constraints
  - lexicographic-optimization
  - benchmark
pretty_name: SoftOR
configs:
  - config_name: default
    default: true
    data_files:
      - split: test
        path: data/*.parquet
  - config_name: compass
    data_files:
      - split: test
        path: data/compass.parquet
  - config_name: travelplanner
    data_files:
      - split: test
        path: data/travelplanner.parquet
  - config_name: lexibench
    data_files:
      - split: test
        path: data/lexibench.parquet
  - config_name: nurse_rostering
    data_files:
      - split: test
        path: data/nurse_rostering.parquet

SoftOR

SoftOR is a 355-instance benchmark for translating natural-language decision problems into optimization models that distinguish hard requirements from soft preferences. It accompanies ASTOR, a method for compiling grammar-guided abstract syntax trees into solver code.

This release contains the exact instance IDs from the paper's main evaluation cohort, task text from saved experiment records, external instance data, and reference annotations or gold evaluation records. It does not include model-generated answers or experiment logs.

Components

Component	Instances	Evaluation
COMPASS	85 (78 hotel, 7 complex)	Hard feasibility and equal-priority preference satisfaction, using the source oracle
TravelPlanner	180	Hard feasibility, ordered preference satisfaction where present, and cost
LexiBench	75	Hard feasibility, lexicographic feature counts, and an optional final tie-break
Nurse Rostering	15	Hard feasibility and published-weight staffing/request penalties
Total	355

All instances are in a single test split. This is an evaluation split, not a new random train/test partition. TravelPlanner instances come from its upstream validation set. Of the 180 TravelPlanner instances, 95 have soft-preference tiers and 85 are cost-minimization instances without such tiers. COMPASS here is the 85-instance cohort used in the paper, not the separate experimental compass_lexicographic or compass_lexicompose datasets.

Load the dataset

from datasets import load_dataset

ds = load_dataset("amergoel/SoftOR", split="test")
lexibench = load_dataset("amergoel/SoftOR", "lexibench", split="test")
print(ds[0]["problem_text"])

The table keeps large structured inputs and evaluation artifacts in separate compressed JSON files. Download a row's input like this:

import gzip
import json
from huggingface_hub import hf_hub_download

row = ds[0]
path = hf_hub_download(
    "amergoel/SoftOR", row["input_path"], repo_type="dataset"
)
with gzip.open(path, "rt", encoding="utf-8") as f:
    instance = json.load(f)

problem_text = instance["problem_text"]
external_parameters = instance["parameters"]
parameter_schema = instance["parameter_schema"]

For local evaluation, download reference_path in the same way. Reference files contain gold answers and evaluator annotations. Do not include them in model prompts.

In the paper's blind formulation setup, the model receives the problem text and structural data contract. Numerical external parameters are bound after formulation. Direct-answer methods need access to the instance data to produce a decision. Supplying all numeric data to every formulation method is a different protocol. The problem_text field includes the recorded data contract, but not each method's full system prompt or benchmark adapter instructions.

Fields

Field	Meaning
`id`	Stable instance identifier
`benchmark`	`compass`, `travelplanner`, `lexibench`, or `nurse_rostering`
`subset`	Source stratum, distinguishing hotel and complex COMPASS
`problem_text`	Recorded natural-language task plus structural data contract
`objective_type`	Weighted penalty, lexicographic tiers, equal-priority feature count, or cost minimization
`difficulty`, `problem_type`, `priority_signal`	Reporting metadata from the paper cohort. Not additional model instructions.
`input_path`	Repository path to compressed JSON with the task and external parameter data
`reference_path`	Repository path to compressed evaluator-only JSON

Reference files

COMPASS: original structured task, exhaustive feasible-selection oracle with utility scores, and the mapping from candidate IDs to booking components. Higher oracle utility is better. Evaluate against the supplied oracle, not a newly inferred replacement objective.
TravelPlanner: original task record, saved gold record, lexicographic gold when available, and the constructed candidate bundle. Preference tiers are compared in order before the cost tie-break. Source constraints and feasibility checks still apply.
LexiBench: the original self-contained benchmark record, including task, annotation, and evaluation. Tiers count satisfied preferences, highest priority first. The gold record includes a witness and solver certificate.
Nurse Rostering: source row, canonical scheduling instance, and published gold objective. The main protocol hides published numerical request weights from modeling and uses them only for evaluation. Methods infer their own weights from the preference wording. A hard-infeasible roster is not a correct solution regardless of its penalty.

This repository distributes evaluation data, not a new universal scoring function. Reproducing the paper requires the benchmark-specific feasibility and scoring routines and method prompts from the ASTOR/SoftOR code artifact. In particular, matching an objective number alone does not establish feasibility.

Construction and provenance

COMPASS: natural-language adaptations of the publicly released COMPASS benchmark. Source booking data and oracle semantics are retained.
TravelPlanner: tasks and reference information from OSU NLP's TravelPlanner, with SoftOR preference interpretation and gold evaluation artifacts. See the source paper.
LexiBench: 75 original instances covering three problem types, five preference patterns, and five domains per combination. Gold lexicographic values were certified during the original benchmark construction. This upload does not rerun those solver certifications.
Nurse Rostering: natural-language adaptations of SchedulingBenchmarks Nurse Rostering, using its published scheduling data and reference objectives. See the source citation.

The dataset was prepared using author-designed construction and annotation procedures, including generative-AI assistance for natural-language rendering as described in the paper. Scenarios concern travel planning and staff scheduling. This is a research benchmark, not a validated operational scheduling or booking service.

manifest.json records the cohort and provenance. SHA256SUMS records release-file checksums. Task text is taken from saved main-run records. External parameter data is reconstructed using the existing blind-mode adapters, except Nurse Rostering inputs come directly from the saved public-input snapshots. Reference records are taken from the existing local benchmark artifacts. No new model experiments were run for this release.

License and attribution

Original SoftOR contributions are released under CC BY 4.0. Upstream material retains its own terms. COMPASS's MIT notice and TravelPlanner's code MIT notice are preserved in licenses/. The TravelPlanner dataset is CC BY 4.0. The SchedulingBenchmarks NRP page did not state an explicit redistribution license when checked, so this release does not label that upstream material as CC BY or MIT. See LICENSE.md and THIRD_PARTY_NOTICES.md for scope and attribution.

Please credit SoftOR/ASTOR and the applicable source benchmarks when using the data. A final paper citation can be added when its public bibliographic identifier is available.