README.md

metadata

license: cc-by-4.0
pretty_name: TurnFlow
language:
  - en
  - zh
task_categories:
  - text-generation
tags:
  - benchmark
  - evaluation
  - multi-turn
  - user-simulation
size_categories:
  - n<1K
configs:
  - config_name: scenarios
    default: true
    data_files:
      - split: test
        path: scenarios/*.jsonl
  - config_name: judges
    data_files:
      - split: test
        path: judges/*.jsonl

TurnFlow

A versioned collection of 490 synthetic multi-turn scenarios and their source-bound judging contracts. The corresponding simulator, runner, compiler, and scoring rules are available at X-EASys/TurnFlow. Version 0.4.0 uses contiguous task IDs T1–T5, based on the rc5 source snapshot. This version changes task/source identifiers and derived hashes; scenario content and scoring rules are unchanged. Earlier tags retain their original identifiers.

Task	Description	Sources
T1	Stateful Multi-Constraint Following	100
T2	Scope Creep	79
T3	Constraint Collision	100
T4	Preference Override	97
T5	Goal Shift	114

Load the data

import json
from datasets import load_dataset

sources = load_dataset("Buka7yo/TurnFlow", "scenarios", revision="v0.4.0", split="test")
judges = load_dataset("Buka7yo/TurnFlow", "judges", revision="v0.4.0", split="test")
scenario = json.loads(sources[0]["scenario_json"])
contract = json.loads(judges[0]["definition_json"])

Both configurations contain the same 490 source keys in a single test split. There is no training or validation split. T1–T5 identify the five released tasks; the full source_key distinguishes variants sharing a bare ID.

Scenario rows

• source_key: complete source identity, for example T4/v3/T4_400.
• task: T1, T2, T3, T4, or T5.
• domain: the source's scenario domain.
• source_sha256: SHA256 of the UTF-8 bytes in scenario_json.
• scenario_json: the exact source JSON text, preserving the packaged source bytes.

The parsed scenario contains the initial situation, private state/aim sequence, constraints, and ground-truth information. The evaluated assistant must receive only the conversation and public base instruction, without this private state or the judge definition. The released runner enforces this separation.

Judge rows

• source_key, task, source_sha256: join keys back to the scenario.
• profile: source-specific parser/representation profile.
• definition_json: the frozen judge definition; T2/T4 include the source-bound template and structured specification, while T1/T3/T5 use their fact definitions.
• definition_sha256: SHA256 of the UTF-8 definition_json string.
• catalog_json: source-binding and definition metadata used by the matching runtime.

These contracts depend on the matching compiler and parser. They are not a list of scalar labels for completed assistant responses. Generic judge prompting or changing aggregation thresholds defines a different evaluation protocol.

Additional files

protocol.json specifies the simulator/judge defaults and source ordering. bindings.json selects source-bound simulator rendering. runtime_resources/ contains the JSON resources used by the code's prompt compiler and reducer. manifest.json records per-source, per-definition, and file hashes, plus the matching code fingerprint and dataset-manifest hash.

For reproduction, pin the GitHub tag v0.4.0 and this repository's v0.4.0 tag. The GitHub release.json also pins the exact Hugging Face commit. Installation of the Python package includes all sources and needs no Hugging Face account.

Intended use and limitations

This dataset supports evaluating assistants' behavior across evolving goals, constraints, scope changes, and preferences in multi-turn conversations. The scenarios are constructed evaluation material, rather than real users' chat logs. All sources and judging contracts are disclosed for reproducibility; there is no claim that they are unseen by model training.

The default protocol uses Gemini 3.7 Flash for simulation and DeepSeek V4 Flash with high reasoning for a single judge call per completed dialogue. Fact extraction is probabilistic even though aggregation is deterministic. Source ambiguities, interpretation boundaries, provider failures and model changes can affect scores. This release does not establish universal judge correctness or invariant model rankings. Report all model settings, dataset/code revisions, planned denominators, coverage, and separate PASS/FAIL/U/NA/INVALID counts.

The repository contains benchmark inputs and contracts. Generated dialogues, model performance tables, experiment logs, billing records, and internal analysis reports are not distributed with it. The software and task documentation are at GitHub.

License

The dataset materials are licensed under CC BY 4.0. The matching Python software is licensed separately under MIT. Please credit TurnFlow and its contributors, link to the license, and indicate changes when sharing adapted materials.