SWE-Rebench-Tasks-Clean

1,322 verified-solvable, contamination-controlled software-engineering tasks for terminal-agent RL training.

Adapted from nebius/SWE-rebench-V2 (real GitHub issue → PR tasks with executable test contracts) into the TerminalWorld task format. Companion dataset to Fzz1/SWE-Smith-Seeds-Clean, same layout.

Every task is a directory containing:

file	content
instruction.md	the issue text the agent sees (plus linked issue discussion where available)
environment/	Dockerfile pointing at the upstream per-task image, frozen at the pre-fix base commit
solution/	gold fix patch + solve.sh reference solution
tests/	test.sh grader + test.patch (FAIL_TO_PASS tests + sampled PASS_TO_PASS regression anchor)
task.toml	resources: 600 s grader budget, 1 CPU, 4 GB
seed.json	full provenance: repo, base commit, test contract, quality metadata

How tasks were selected

Source pool: SWE-rebench-V2 train split (~21k tasks), annotated per-task with tb_category, difficulty, quality_code (A / B1–B6), intent_completeness (complete / partial / insufficient), and terminal_shaped (strong / some / library_patch_only).

The dataset is three batches, kept separate as the tier column. Quality differs by tier on purpose — weight or filter by tier rather than merging blindly.

tier	selection rule
batch1	original category-quota draw (scarce categories taken in full)
batch2	reserve-pool premium: quality A + complete intent + terminal workflow; scarce categories (games, video, ML, model-training, security, optimization) taken in full regardless of grade
batch3	remaining complete-intent reserve; mostly library_patch_only, B-grades admitted; lowest tier

Filtering funnel

Per batch: how many entered each stage and how many survived it.

stage	what it does	batch1	batch2	batch3	total
selected	metadata screen on the upstream pool (categories, quality, intent)	1,100	1,282	755	3,137
adapted	task generation succeeds (drops: malformed rows, dead upstream images)	1,091	1,236	749	3,076
oracle ship	gold fix earns reward 1 at training conditions (Gate 1)	931	1,052	663	2,646
kept	solver earns reward 1 at least once (Gate 2)	448	535	339	1,322

Stage survival rates — oracle: 85% / 85% / 89% of adapted; solve: 48% / 51% / 51% of shipped. batch1's kept set is 390 tasks solved by gpt-5.6-sol (effort high, single attempt) plus 58 recovered by gpt-5.6-luna (effort max, n=5) from the sol-unsolved remainder; batch2/3 used the luna n=5 gate throughout.

Two-stage filtering

Every task passed both gates:

Gate 1 — oracle solvability. The gold fix must earn reward 1 under exact training conditions (600 s, 1 CPU / 4 GB, network blocked): all FAIL_TO_PASS tests pass and a ≤200-test sampled PASS_TO_PASS regression anchor stays green. This proves the task is solvable as shipped. Oracle pass rates: 85% (batch2), 89% (batch3) of adapted candidates.

Gate 2 — on-policy solvability. gpt-5.6-luna (reasoning effort max) attempts each task 5 times; the task is kept only if at least one attempt scores reward 1. Tasks the model never solves give no learning signal and were dropped. batch1 additionally contains 390 tasks solved by gpt-5.6-sol (effort high) in one attempt.

The parquet records solver and passrate per task. Passrate is a usable difficulty signal: 5/5 = solved every attempt (easy), 1/5 = solved once in five (hardest kept).

Contamination controls

Four answer-leak channels were discovered during solving and closed. Each closure was verified behaviorally (by the leak not reproducing on its worst known case), not just by configuration:

1. Hosted web search — executed provider-side, invisible to sandbox egress blocks; agents were fetching the exact fix PRs. Disabled; verified zero search events on a task that previously searched 9 times. used_web_search is recorded on every solving record.
2. Git history — images carried the fix commit; agents ran git log --all and read it. Now scrubbed pre-agent: detach at base commit, delete all refs, expire reflog, prune; asserted that the reachable set equals HEAD ancestry and fsck --unreachable is empty.
3. Newer same-project source in the image — site-packages and sibling conda environments (e.g. an evaluation env) contained the post-fix library; agents diffed their repo against it. Every external copy of the project's own modules is now neutralized (symlinked back to the repo) before the agent starts; the run fails closed if a copy remains.
4. API key exposure — the key was visible on the process command line, then in the inherited environment. Moved to codex's auth store; verified absent from trajectories including projects whose own test suites echo the environment.

Additionally, a 45-task pilot was audited trajectory-by-trajectory: zero reward hacks (no grader tampering, no test edits, no reward-file writes). Contract errors on the failure side (mis-attributed FAIL_TO_PASS tests) were repaired corpus-wide before the final runs.

Distribution

By tier:

tier	tasks
batch2	535
batch1	448
batch3	339

By solver and passrate:

solver	passrate	tasks
gpt-5.6-luna (max, n=5)	5/5	694
gpt-5.6-sol (high, n=1)	1/1	390
gpt-5.6-luna (max, n=5)	4/5	69
gpt-5.6-luna (max, n=5)	3/5	41
gpt-5.6-luna (max, n=5)	2/5	54
gpt-5.6-luna (max, n=5)	1/5	74

By category:

category	tasks		category	tasks
debugging	273		security	66
system-administration	178		data-processing	46
data-querying	174		scientific-computing	45
mathematics	168		machine-learning	43
software-engineering	106		model-training	41
file-operations	96		optimization	38
			data-science	28
			games	18
			video-processing	2

Files

path	content
data/tasks-00000.tar	all 1,322 task directories under tasks/<instance_id>/
metadata/tasks.parquet	one row per task: instance_id, tier, solver, passrate, tb_category, repo, image, base_commit, difficulty, n_f2p, n_p2p_sampled, quality_code, intent_completeness, terminal_shaped, thin_anchor, contract_unresolved
metadata/task_ids.txt	all instance ids, one per line
metadata/batch1_ids.txt etc.	per-tier id lists
metadata/images.txt	all Docker images referenced
metadata/validation_summary.json	pipeline description + counts
metadata/passrate_distribution.json	passrate histogram per tier and solver

Caveats

• tb_category describes the upstream project's domain, not the task's work. It is assigned largely from repo identity — 64% of repos have all their tasks under a single category — so a one-line constant change in conda carries system-administration. Do not read the category table as coverage of Terminal-Bench's categories. An audit against Terminal-Bench 2 found that all 89 of its tasks grade a produced artifact (tests/test_outputs.py reading a file the agent created) while every task here grades the upstream project's own pytest after a source patch; the two are different task shapes that happen to share category names. These tasks are software-engineering data — reproduce from an issue, navigate an unfamiliar repo, write a minimal patch — and should be valued as that.
• thin_anchor marks tasks whose PASS_TO_PASS anchor is small (fewer than ~10 tests); a fix could break the repo without the grader noticing. Recorded, not repaired.
• contract_unresolved flags tasks with an unresolved FAIL_TO_PASS attribution suspicion; they are graded on their full contracts.
• Reward-0 rows from the solving runs are not in this dataset, deliberately. A sampled deep audit found that a large fraction of solve failures trace to instruction underspecification — the tests pin exact strings, names, or counts the issue text never states. Do not mine "the model failed this task" populations from this corpus as clean training negatives without screening for that.