gradients-io-tournaments/SWE-ZERO-12M-trajectories-filtered

Filtered SWE Trajectories — 60K

This dataset contains 60,000 software-engineering agent trajectories from two upstream datasets:

• —50,000 Submitted trajectories sampled from `AlienKevin/SWE-ZERO-12M-trajectories`
• —10,000 accepted trajectories sampled from `Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k`

All rows use this common schema:

• —instance_id
• —repo
• —messages
• —trajectory_format
• —exit_status
• —duration_sec

For the SWE-smith rows, repo was derived from the source instance_id, trajectory_format is mini-swe-agent-plus, and exit_status is Accepted. The upstream SWE-smith schema does not publish wall-clock duration, so duration_sec is null for those 10,000 rows rather than an invented value.

Sampling

SWE-ZERO 50K

The first subset was constructed from upstream revision 44e028077c55e7255c328516c8bd76080fbb3840 with deterministic seed 20260710:

1. 1.180 source Parquet shards were selected at random from the complete 1,230-shard manifest.
2. 2.Only rows whose upstream exit_status was exactly Submitted were eligible.
3. 3.Eligible rows received deterministic SHA-256 random priorities.
4. 4.The 50,000 eligible rows with the best priorities were retained.

The sampling is intentionally not claimed to be a uniform sample of every row in the 12.3M-row upstream dataset. It is a randomized, diverse subset spanning 7,398 unique pull-request tasks and 1,959 repositories.

SWE-smith 10K

The additional subset was constructed from upstream revision 750b2c11239fd5e32f97e6cfb9bf80fb9a9a2983 with deterministic seed 20260806:

1. 1.10 source Parquet shards were selected at random from the 47-shard manifest.
2. 2.All 14,042 rows in those shards were transformed to the common schema.
3. 3.Rows received deterministic SHA-256 random priorities.
4. 4.The 10,000 rows with the best priorities were retained.

These rows span 6,648 unique SWE-smith task IDs and 123 repositories. Multiple accepted trajectories may correspond to the same task ID.

Decontamination

For both subsets, every candidate was checked against a private, vetted 74-task SWE-INFINITE evaluation set. A row was ineligible if its instance_id exactly matched a protected task. The final combined JSONL was independently checked for protected-set overlap. The protected task identities are deliberately not published with this training dataset.

Validation

• —Rows: 60,000
• —exit_status == "Submitted" rows: 50,000
• —exit_status == "Accepted" rows: 10,000
• —trajectory_format == "mini-swe-agent-1" rows: 50,000
• —trajectory_format == "mini-swe-agent-plus" rows: 10,000
• —Rows with duration_sec == null: 10,000
• —Unique task IDs: 14,046
• —Unique repositories: 2,059
• —Exact protected instance_id overlaps: 0
• —SHA-256 of data/train.jsonl: b4439b9646f20e893cee15c9c4ab4b57632947d0d2cd80744201b5e884d29491

License and attribution

This is a combined dataset with mixed upstream licensing. The SWE-ZERO source is distributed under Apache-2.0, while the SWE-smith source is distributed under MIT. Users should review both upstream dataset cards for their complete methodology, limitations, attribution, and license terms.