MetaCurate-DeepSeekV4Pro

An uncurated corpus of 4,286 coding-agent trajectories covering 1,415 instance IDs, generated with DeepSeek V4 Pro and mini-swe-agent for MetaCurate: Evolving Curation Filters for High-Quality Agentic Coding Data.

Project page · Code

This is a public test release under the maintainer's account. The intended official repository is furiosa-ai/MetaCurate-DeepSeekV4Pro.

The full pool includes successful, unsuccessful, and indeterminate attempts. It is the paper's DSv4Pro pool, before MetaCurate or baseline subset selection. No trajectory was removed or truncated during packaging. Reasoning, tool calls, tool observations, final patches, and original harness metadata are retained.

Load the data

from datasets import load_dataset

data = load_dataset("JakeOh/MetaCurate-DeepSeekV4Pro", split="train")
print(len(data))  # 4286
example = data[0]

For reproducible use, pass a release commit SHA with revision=.... All rows belong to the single train split; no new random train/test split is introduced.

Columns

Column	Type	Meaning
id	string	Unique release ID, such as dsv4pro-000000. Use this key for joins and subset lists.
messages	list of structs	Ordered original conversation and harness events.
instance_id	string	Upstream issue/instance ID; multiple attempts can share it.
repo	string	Target GitHub repository.
license	string	License annotation from the source data. It may describe the dataset or repository and may contain unknown or custom-check-github; it is not an independently verified SPDX assertion.
trajectory_id	string	Original experiment identifier, retained for traceability. It is not unique.
dataset	string	Original collection source: nebius/SWE-rebench, nebius/SWE-rebench-V2, or SWE-Gym/SWE-Gym.
model_patch	string	Original final submitted patch, including empty patches.
base_commit	string	Repository commit reconstructed from the identified collection source and upstream instance metadata.
resolved	nullable boolean	Archival grading outcome under the conservative policy below: supported pass, supported failure, or indeterminate.

messages preserves content, reasoning_content, tool_calls, tool_call_id, provider fields, and harness extra fields. Function-call arguments remain JSON strings. Message roles include system, user, assistant, tool, and the harness-specific terminal exit event. Consumers should apply their own chat-template and loss-masking policy; do not assume every event is a standard chat role. Existing terminal events and missing final tool responses are not synthetically repaired.

The common bash function schema is provided in tools.json. Arrow represents absent struct fields as null; original non-null values and sequence order are preserved.

Collection and provenance

The recorded generator model is deepseek/deepseek-v4-pro-20260423. Collection used mini-swe-agent at temperature 0.7, targeting three attempts per source instance. Overlapping source issues and repeated attempts were retained.

Collection source	Source-specific instance IDs	Trajectories
SWE-rebench	690	2,100
SWE-rebench-V2	531	1,599
SWE-Gym	196	587
Combined	1,415	4,286

Source-specific instance counts overlap and should not be summed to infer the combined unique count.

The frozen source JSONL has 3,267,571,571 bytes and SHA-256:

304c353e9461b4d8d06c22348bbc69c7ada44ccad9c2f31ae795850fe77cd9cd

Release IDs are assigned once from the zero-based row index of that exact JSONL, padded to six digits. Its original manifest uses a different order, so manifest position must not be used to reconstruct IDs. The pool contains 4,280 unique legacy trajectory_id values and six extra occurrences with different trajectory content; all 4,286 rows are retained.

The old SFT export misattributed 69 rows to SWE-Gym and carried the associated MIT annotation. Original rollout paths, with content-and-patch matching for colliding identifiers, identify 24 of these as SWE-rebench and 45 as SWE-rebench-V2. This release corrects only their dataset and license annotations, in addition to adding the three release columns. The exact old/new values and source-row hashes are in provenance/source-annotation-corrections.jsonl. Original messages, patches, instance IDs, repositories, and legacy trajectory IDs are unchanged.

provenance/row-annotations.parquet records the source-row hash and grading-label status for each id. provenance/release.json records source revisions, transformation rules, and output checksums.

Grading labels and limitations

The release uses surviving collection-time grading ledgers. It does not rerun grading or replace missing evidence with guessed outcomes.

resolved	Rows	Interpretation
true	1,758	Unambiguous ledger records report both fail-to-pass and pass-to-pass checks passing.
false	1,671	An unambiguous record reports a failed check and a non-vacuously passed check.
null	857	The surviving evidence cannot support either label under this policy.

The null rows comprise:

• 464 records where both checks are false and no detailed report survives to distinguish a failed solution from a verification error.
• 391 records where fail-to-pass is false and pass-to-pass is true only because the source has no pass-to-pass tests.
• Two different trajectories for dask__dask-7145.local.1 whose duplicate ledger key has conflicting pass/fail records without enough attribution to assign the outcomes.

The historical source README and grading ledger report 1,759 passes; the manuscript reports 1,760. This packaging audit cannot substantiate the manuscript count. The conservative release has 1,758 unambiguously attributable passes and leaves the conflicting pair unlabeled. Do not convert null to false or treat these reconstructed counts as a new benchmark evaluation.

Detailed grading sidecars are unavailable in the surviving ledger export. Labels inherit the original grader's test definitions and limitations. The paper describes overlap checks against SWE-bench Verified; this packaging release does not establish protection against every form of semantic benchmark contamination.

License and attribution

FuriosaAI's original dataset contributions and compilation are made available under CC BY 4.0. Source data, issue text, code, patches, and third-party material retain their applicable licenses; this grant does not replace those licenses. See NOTICE.md.

SWE-rebench and SWE-rebench-V2 are distributed under CC BY 4.0 and require respecting the licenses of their source repositories. SWE-Gym is distributed under MIT. The per-row license annotation is a provenance hint; consult the referenced upstream dataset and repository for the applicable terms.

Citation

@misc{kim2026metacurate,
  title = {MetaCurate: Evolving Curation Filters for High-Quality Agentic Coding Data},
  author = {Kim, Donghoon and Oh, Seunghyuk and Galim, Kevin and Kim, Minseo and Lee, Minjae and Kim, Younghyun and Seo, Wooseok and Koo, Hyung Il and Kang, Wonjun},
  year = {2026},
  url = {https://furiosa-air.github.io/metacurate/}
}