JetBrains-Research/SWE-TaskFlow-SWE-Bench-Pro

SWE-TaskFlow: SWE-Bench Pro Interactive Task Flows

This is the unified data release for “Opinion: Coding-Agent Benchmarks Should Match Their Users’ Task Flows” by Igor Slinko, Yaroslav Golubev, and Sergey Titov.

SWE-TaskFlow transforms verified SWE-Bench Pro tasks into replayable interactive trajectories while preserving the original repositories, requirements, and test-based verification. The dataset contains:

• issue-solving tasks presented as one prompt or as a sequence of actionable prompts;
• repository questions inserted at controlled positions in those sequences;
• hidden reference answers and executable proof artifacts for evaluating QA turns;
• the TFAS-optimized task-flow selection used as the paper’s main Split+QA condition.

All experimental conditions are exposed as splits of one dataset. There are no separate Hugging Face configurations to select.

Evaluation leakage warning: QA splits contain reference answers, accepted and incorrect answers, evidence, proof scripts, and proof-execution reports. Give the evaluated agent only the ordered prompt texts from hybrid. Keep all other QA fields hidden from the agent.

Splits

Issue-solving conditions

Split	Paper condition	Rows	Agent interaction
single_prompt	Single	700	Original SWE-Bench Pro task delivered as one user turn
split_prompt	Split	700	The same task decomposed into three ordered user turns
split_prompt_concat	Concat	700	The three split segments concatenated into one turn

These splits contain the same ordered 700 SWE-Bench Pro instances. One otherwise eligible Teleport task is excluded because its repository context exceeds the evaluation context window. Single versus Concat measures rewriting effects; Concat versus Split isolates sequential disclosure.

QA-augmented conditions

Split	Rows	Questions	Purpose
qa_tfas_optimized_sol	700	454	Main TFAS-optimized Split+QA benchmark; questions occur in 255 eligible sessions
qa_first_sol	255	467	Sol questions before issue turns
qa_last_sol	255	467	Sol questions after issue turns
qa_random_sol	255	467	Sol questions at randomized positions
qa_only_sol	255	467	Sol question turns without issue turns
qa_first_terra	64	114	Terra placement ablation: questions first
qa_last_terra	64	114	Terra placement ablation: questions last
qa_random_terra	64	114	Terra placement ablation: randomized positions
qa_only_terra	67	118	Terra question turns without issue turns

Within each generator family, placement splits reuse the same question payloads and differ only in prompt ordering. This supports paired comparisons by instance_id.

Quickstart

import json

from datasets import load_dataset


dataset_id = "JetBrains-Research/SWE-TaskFlow-SWE-Bench-Pro"
split = "qa_tfas_optimized_sol"
dataset = load_dataset(dataset_id, split=split)

row = dataset[0]
trajectory = json.loads(row["hybrid"])

for prompt in trajectory["prompts"]:
    print(prompt["kind"], prompt["text"])

Load another experimental condition by changing only split:

single = load_dataset(dataset_id, split="single_prompt")
split = load_dataset(dataset_id, split="split_prompt")
concat = load_dataset(dataset_id, split="split_prompt_concat")

What the agent sees

Parse hybrid as JSON and deliver its prompts sequentially within one conversation:

{
  "schema_version": "qa_multistep_prompt/v1",
  "instance_id": "instance_ansible__ansible-...",
  "prompts": [
    {"kind": "issue", "part_index": 0, "text": "..."},
    {"kind": "qa", "question_id": "q-1", "text": "..."},
    {"kind": "issue", "part_index": 1, "text": "..."}
  ],
  "record_sha256": "..."
}

During issue turns, the agent may inspect, edit, and test the repository. During QA turns, the agent answers in free text; QA answers should not be inserted into the repository patch.

Unified row schema

Every split uses the same Parquet schema:

Column	Description
instance_id	SWE-Bench Pro instance ID and cross-split join key
problem_statement	Original SWE-Bench Pro issue description
subset	Source subset identifier
phase_order_config	Original condition identifier; legacy values use hyphens while Hub split names use underscores
stable_question_ids	Ordered question IDs; null for issue-only splits
schema_version	Row schema version
record_sha256	Content hash of the canonical hybrid record
bundle_sha256	Content hash of the validated QA bundle; null for issue-only splits
hybrid	JSON-encoded ordered interactive trajectory
questions	JSON-encoded hidden QA records; null for issue-only splits
golden_proof_scripts	JSON map from question IDs to executable proof scripts
golden_proof_executions	Creation-time proof-execution reports
facts	JSON-encoded repository facts used to construct questions
documentation	Human-readable QA bundle documentation
bundle_metadata	JSON-encoded QA bundle metadata

The QA-only columns are nullable so that issue-only and QA-augmented conditions can be represented as splits in one dataset without changing their payloads.

Evaluation protocol

1. Obtain the repository checkout, environment, and verifier tests for each instance_id from the original SWE-Bench Pro release.
2. Parse hybrid and send each prompt as the next user message in the same agent session.
3. Run the source benchmark’s tests/test.sh on the cumulative patch after the final issue turn. For checkpoint signals, snapshot and test after each issue turn against requirements revealed up to that point.
4. For QA turns, judge the agent’s answer against the hidden reference answer and evidence. The paper uses an LLM judge with the reference answer and evidence chain.

This release contains prompts and validation metadata. It does not redistribute repositories, gold patches, SWE-Bench Pro verifier tests, or agent trajectories.

Proof-carrying QA bundles

QA rows include:

• the question and hidden golden answer;
• accepted and incorrect answer examples;
• an answer contract and evidence chain;
• an executable golden proof script;
• a creation-time execution report confirming that the proof ran successfully in a clean repository checkout.

To materialize a QA row as an on-disk bundle and validate its hashes and cross-references:

python scripts/materialize_bundle.py --row-json row.json --output bundles/

Provenance and integrity

• manifest/unified.json records the source dataset revisions, split row counts, Parquet hashes, the schema-normalization step used to combine the releases, and the excluded task.
• manifest/issue-provenance.json records prompt-splitting inputs, generation settings, validation results, and ordered instance IDs.
• manifest/issue-qa-provenance.json records question generation, filtering, proof execution, TFAS optimization, and stable question mappings.
• record_sha256 content-addresses the canonical hybrid record.
• bundle_sha256 content-addresses the full validated QA bundle.
• JSON Schemas are available under schema/.

The unified release changes the Hub layout, adds nullable QA columns to issue-only Parquet rows, and excludes one Teleport task whose repository context exceeds the evaluation context window. Existing trajectory and bundle hashes for retained rows are preserved.

License

No additional license is granted by this dataset card. Prompt and question text is derived from public SWE-Bench Pro tasks and source repositories. Repository and task content remain subject to their respective upstream licenses.

Citation

@inproceedings{slinko2026swetaskflow,
  title={Opinion: Coding-Agent Benchmarks Should Match Their Users' Task Flows},
  author={Slinko, Igor and Golubev, Yaroslav and Titov, Sergey},
  booktitle={NeurIPS 2026 Workshop on Evaluation of Interactive Agents},
  year={2026}
}