Datasets:

f-labs-io
/

SpecSuite-Core

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - benchmark
  - agents
  - llm-agents
  - test-driven-development
  - specification
  - mutation-testing
pretty_name: SpecSuite-Core
arxiv: 2603.08806
size_categories:
  - n<1K
configs:
  - config_name: specs
    data_files:
      - split: train
        path: data/specs/train.jsonl
  - config_name: mutations
    data_files:
      - split: train
        path: data/mutations/train.jsonl
  - config_name: results
    data_files:
      - split: train
        path: data/results/train.jsonl

SpecSuite-Core

SpecSuite-Core is a benchmark suite of 4 deeply-specified AI agent behavioral specifications, designed to evaluate the TDAD (Test-Driven Agent Definition) methodology described in our paper.

Paper: Test-Driven Agent Definition (TDAD) Code: GitHub

Overview

Each specification defines a complete agent behavior contract including:

Tools with typed input/output schemas and failure modes
Policies with priorities and enforcement levels
Decision trees defining the agent's control flow
Response contracts with required JSON fields
Mutation intents for robustness testing
Spec evolution (v1 → v2) for backward compatibility testing

Specifications

Spec	Domain	v1 Tools	v1 Policies	v2 Change
SupportOps	Customer support (cancel, address, billing)	7	4	Abuse detection
DataInsights	SQL analytics & reporting	3	4	Cost-aware queries
IncidentRunbook	Incident response & escalation	6	4	Customer impact tracking
ExpenseGuard	Expense approval & reimbursement	6	5	Manager approval gate

Configs

`specs` — Agent Specifications (8 rows)

Each row is one spec version (4 specs × 2 versions). The spec_yaml field contains the full YAML specification.

from datasets import load_dataset
specs = load_dataset("f-labs-io/SpecSuite-Core", "specs")

# Get SupportOps v1
spec = specs["train"].filter(lambda x: x["spec_id"] == "supportops_v1")[0]
print(spec["title"])        # "SupportOps Agent"
print(spec["tool_names"])   # ["verify_identity", "get_account", ...]
print(spec["spec_yaml"])    # Full YAML specification

`mutations` — Mutation Intents (27 rows)

Each row is a mutation intent: a description of a plausible behavioral failure that the test suite should detect.

mutations = load_dataset("f-labs-io/SpecSuite-Core", "mutations")

# All critical mutations
critical = mutations["train"].filter(lambda x: x["severity"] == "critical")
for m in critical:
    print(f"{m['spec_lineage']}/{m['mutation_id']}: {m['intent'][:80]}...")

Categories: policy_violation, process_violation, business_logic_violation, grounding_violation, escalation_violation, safety_violation, quality_regression, compliance_violation, robustness_violation, decision_violation, tooling_violation

`results` — Pipeline Run Results (24 rows)

Each row is one end-to-end TDAD pipeline run with metrics across all 4 evaluation dimensions.

results = load_dataset("f-labs-io/SpecSuite-Core", "results")

for r in results["train"]:
    print(f"{r['spec']}/{r['version']}: VPR={r['vpr_percent']}% HPR={r['hpr_percent']}% MS={r['mutation_score']}% ${r['total_cost_usd']:.2f}")

Metrics

Metric	Full Name	What It Measures
VPR	Visible Pass Rate	Compilation success (tests seen during prompt optimization)
HPR	Hidden Pass Rate	Generalization (held-out tests never seen during compilation)
MS	Mutation Score	Robustness (% of seeded behavioral faults detected by tests)
SURS	Spec Update Regression Score	Backward compatibility (v1 tests passing on v2 prompt)

Full Pipeline

The specifications in this dataset are designed to be used with the TDAD pipeline:

TestSmith generates executable tests from the spec
PromptSmith iteratively compiles a system prompt until tests pass
MutationSmith generates behavioral mutations to measure test quality
Tests and mutations measure the agent across VPR, HPR, MS, and SURS

See the GitHub repo for the full executable pipeline with Docker support.

Citation

@article{rehan2026tdad,
  title={Test-Driven Agent Definition: A Specification-First Framework for LLM Agent Development},
  author={Rehan, Tzafrir},
  year={2026}
}

License

Apache 2.0