CoEvoke: Multi-Agent Collaboration Trajectories

CoEvoke provides 11,815 structured collaboration episodes derived from 4,567 parent trajectories over 4,130 canonical tasks. It covers software engineering, mathematical reasoning, natural-language logical reasoning, and logic-grid puzzles. Records preserve task inputs, agent outputs, tool interactions, delivered messages, local review decisions, and provenance.

This expanded release reuses the collected trajectories. Each episode contains one complete adjudicated collaboration loop or two adjacent loops, together with the preceding exported public history. Expansion and publication packaging used no new model generation or evaluator calls. The original parent trajectories are not counted again as extra rows.

中文说明：本数据集包含 11,815 条多智能体协作片段，来源于 4,567 条母轨迹、4,130 个去重任务。每条记录保留单轮或相邻双轮协作及其前序公开上下文。条目之间可能共享母轨迹与上下文，因此条目数不等于独立任务数。

Distribution

Source family	Parent trajectories	Canonical tasks	Single-loop episodes	Adjacent double-loop episodes	Total episodes
SWE (swe_oracle)	1,763	1,763	3,315	1,552	4,867
DeepMath (deepmath)	923	831	1,656	733	2,389
ProofWriter (proofwriter)	1,392	1,096	2,315	923	3,238
ZebraLogic (zebra)	489	440	905	416	1,321
Total	4,567	4,130	8,191	3,624	11,815

Source family	train	dev_internal	Total
SWE	4,644	223	4,867
DeepMath	2,261	128	2,389
ProofWriter	3,068	170	3,238
ZebraLogic	1,255	66	1,321
Total	11,228	587	11,815

The splits inherit the original canonical-task grouping. No canonical task or split group crosses the two splits. dev_internal is an internal development split, not an independent external benchmark. Upstream source splits, stored in each task's provenance, are separate from this release's splits.

Loading

import json
from datasets import load_dataset

dataset = load_dataset("Some-thy/CoEvoke")
assert len(dataset["train"]) == 11228
assert len(dataset["dev_internal"]) == 587

row = dataset["train"][0]
episode = json.loads(row["episode_json"])
task = episode["task_layer"]
history = episode["context_layer"]
steps = episode["trajectory_layer"]["steps"]
messages = episode["communication_layer"]["messages"]

Streaming is also supported:

stream = load_dataset("Some-thy/CoEvoke", split="train", streaming=True)
episode = json.loads(next(iter(stream))["episode_json"])

The 48 Parquet shards contain every complete episode in the episode_json string column. No custom dataset loader or remote code is required. The JSON string preserves heterogeneous agent/tool payloads without coercing them into a flattened common schema. Separate per-record JSON files are not needed to read the complete records.

Row schema

Column	Type	Meaning
episode_id	string	Unique episode identifier.
parent_trajectory_id	string	Parent trajectory identifier.
canonical_task_id	string	Task identity within the source family.
dataset	string	One of swe_oracle, deepmath, proofwriter, zebra.
split, split_group	string	Inherited split and grouping key.
loop_start, loop_end	int64	Inclusive, one-based review-loop interval.
loop_count, parent_loop_count	int64	Selected loop count and parent loop count.
context_steps, focus_steps	int64	Numbers of preceding-history and focal events.
correctness_status	string	Always NOT_EVALUATED at episode scope.
parent_verdict	string	Parent-level result metadata; not an episode correctness label.
semantic_hash, content_hash	string	Original record-integrity and normalized-content fingerprints.
episode_path	string	Relative locator in the original source package; the Hub copy stores the corresponding full record in episode_json.
episode_json	string	Complete UTF-8 JSON record, including nested content and provenance.

The full JSON Schema is schemas/episode.schema.json. Main JSON sections are:

• task_layer: task input, domain, source identifiers, and recorded source revision where available.
• context_layer: all preceding exported public steps/messages, with independently_executable: false.
• trajectory_layer: run configuration, focal steps, and the local review outcome.
• communication_layer: messages belonging to or entering the focal interval, including their actual delivered payloads.
• boundary: continuation-edge references and whether the real parent endpoint is included.
• provenance: parent identity/hash and selected loop interval.
• parent_terminal_outcome: present only when the selected interval reaches the actual parent endpoint.

Construction and interpretation

A loop ends at a reviewer, critic, or judge output, depending on the task family. A parent with loops A, B, C contributes A, B, C, A+B, and B+C. Prefix context accompanies each selected interval. IDs retain their parent-local meanings and should be resolved within the parent trajectory, not treated as globally unique across all rows.

Steps and messages are preserved from the accepted source release. Tool calls and observations remain paired. An edge may name a recipient after the selected interval, but that recipient's future output is not included. Sparse interactions may use a candidate carried in the preceding history.

The local review decision is not a new ground-truth evaluation. All episode outcomes are NOT_EVALUATED. parent_verdict describes the original whole-task outcome even for an intermediate episode: exclude it from model inputs when that outcome would not yet be known. Only terminal episodes contain parent_terminal_outcome in the JSON payload. Preserve parent outcome categories separately instead of treating unavailable results as failures.

Episodes overlap by construction. Group related examples by canonical task and parent trajectory when creating additional splits or computing uncertainty. Exact normalized-content deduplication removed zero candidates in this release; that does not imply statistical independence or absence of partial overlaps.

Completeness and limits

The release preserves the complete exported public history, not private prompts, hidden reference solutions, Oracle guidance, or repository/environment snapshots. The private reference addon is omitted. The records are therefore suitable for inspecting the exported collaboration process, but do not by themselves guarantee exact execution replay. Tool execution on a downloaded trace is neither required nor performed by the loader.

SWE traces can quote public repository test fixtures, including example passwords and cryptographic test keys. Screening matches in this release were reviewed against the corresponding public test sources and retained as part of the original task context.

Some upstream source_revision values were not recorded and remain null; current upstream revisions are not substituted for unknown historical revisions. Recorded source IDs and task identifiers are retained for attribution and traceability.

Before upload, all 11,815 records were checked against the full JSON Schema and corresponding parent events. Checks also covered manifest membership, unique episode IDs, normalized-content fingerprints, selected-loop coverage, message ordering and references, tool-call pairing, future-output boundaries, split isolation, all source checksums, and exact JSON/Parquet parity. The Parquet files in this repository are byte-for-byte copies of the validated source shards. These checks verify structural integrity and faithful packaging, not the factual correctness of every generated answer.

Files and verification

• data/*.parquet: complete machine-readable episodes.
• schemas/episode.schema.json: complete record schema.
• manifests/episode_manifest.jsonl: original per-episode index and hashes.
• statistics.json: recomputable counts and split distribution.
• provenance.json: source hashes and packaging scope.
• validation_report.json: full-release structural validation results.
• checksums.sha256: SHA-256 digests of the release files, excluding the checksum file itself and Hub-managed files.
• verify_download.py: local checksum, manifest, and payload verification; add --schema to validate every record against JSON Schema.

For a local snapshot:

pip install huggingface_hub pyarrow jsonschema
hf download Some-thy/CoEvoke --repo-type dataset --local-dir CoEvoke
python CoEvoke/verify_download.py --schema

Sources and rights

Task inputs originate from SWE-smith, DeepMath-103K, ProofWriter, and ZebraLogic. The original recorded SWE dataset ID is princeton-nlp/SWE-smith.

Rights are component-specific. This card does not assert a new blanket license covering all upstream tasks, repository code, and model-generated material. See THIRD_PARTY_NOTICES.md for source attribution and the license information present in the upstream dataset cards.