Datasets:

Poindexter-Labs
/

PDL-SWE-Bench

instance_id string	task_id string	task_version int64	tier string	domain string	problem_statement string	patch string	test_patch string	FAIL_TO_PASS list	PASS_TO_PASS list	files list	patch_policy_allowed list	patch_policy_forbidden list	token_budget int64	hidden_test_count int64	public_test_count int64	graded_ladder_rungs list	reward_formula string
pdl_swe_booking_merge_007-v1	pdl_swe_booking_merge_007	1	T3	booking-interval coalescing service	Back-to-back bookings show up as separate schedule blocks instead of one continuous block. For example, a booking from 09:00 to 10:00 followed by another from 10:00 to 11:00 in the same room renders as two entries; it should render as a single 09:00-11:00 block. Bookings that actually overlap merge correctly. Expected ...	--- a/src/merge.py +++ b/src/merge.py @@ -13,7 +13,7 @@ def coalesce(bookings): blocks = [(ordered[0].start, ordered[0].end)] for booking in ordered[1:]: start, end = blocks[-1] - if booking.start < end: + if booking.start <= end: blocks[-1] = (start, max(end, booking.end))...	diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py new file mode 100644 --- /dev/null +++ b/tests/hidden/test_hidden.py @@ -0,0 +1,32 @@ +"""Hidden evaluation tests. Never agent-visible (DEC-015). + +FAIL_TO_PASS: the buggy coalesce treats bookings that meet end-to-end as +disjoint (strict compariso...	[ "tests/hidden/test_hidden.py::test_reported_example_merges", "tests/hidden/test_hidden.py::test_touching_chain_merges_to_one_block", "tests/hidden/test_hidden.py::test_mixed_touching_overlap_and_gaps" ]	[ "tests/public/test_public.py::test_overlapping_bookings_merge", "tests/public/test_public.py::test_disjoint_bookings_stay_separate", "tests/public/test_public.py::test_unsorted_input_is_ordered", "tests/public/test_public.py::test_empty_and_invalid_bookings", "tests/public/test_public.py::test_free_slots_ar...	[ { "content": "# Present so pytest adds the repository root to sys.path, making `import src`\n# resolve from the test suite.\n", "path": "conftest.py" }, { "content": "# Pure standard library unless the environment image pre-bakes wheels.\n", "path": "requirements.lock" }, { "content": ""...	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	250,000	3	6	[ "cheat.diff" ]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_config_merge_002-v1	pdl_swe_config_merge_002	1	T2	layered configuration merging library	When a configuration source overrides one nested setting, every sibling setting in the same section is lost. For example, loading with a config file containing only a database pool size causes the database host, port, and pool timeout to disappear from the effective configuration instead of keeping their defaults. The ...	--- a/src/merge.py +++ b/src/merge.py @@ -9,5 +9,8 @@ """ result = dict(base) for key, value in override.items(): - result[key] = value + if isinstance(value, dict) and isinstance(result.get(key), dict): + result[key] = merge_layers(result[key], value) + else: + ...	diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py new file mode 100644 --- /dev/null +++ b/tests/hidden/test_hidden.py @@ -0,0 +1,33 @@ +"""Hidden evaluation tests. Never agent-visible (DEC-015). + +FAIL_TO_PASS: on the buggy fixture, a partial nested override replaces the +whole section (shallow m...	[ "tests/hidden/test_hidden.py::test_partial_file_override_preserves_siblings", "tests/hidden/test_hidden.py::test_env_override_merges_into_defaults", "tests/hidden/test_hidden.py::test_file_then_env_layers_accumulate" ]	[ "tests/public/test_public.py::test_defaults_when_no_overrides", "tests/public/test_public.py::test_top_level_scalar_override", "tests/public/test_public.py::test_full_section_override", "tests/public/test_public.py::test_env_parsing_builds_nested_tree" ]	[ { "content": "# Present so pytest adds the repository root to sys.path, making `import src`\n# resolve from the test suite.\n", "path": "conftest.py" }, { "content": "# Pure standard library unless the environment image pre-bakes wheels.\n", "path": "requirements.lock" }, { "content": ""...	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	250,000	3	4	[]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_dst_drift_010-v1	pdl_swe_dst_drift_010	1	T3	recurring-event scheduling across timezone transitions	Since the clocks went forward on 30 March, our daily 09:00 stand-up appears at 10:00 in every agenda — it was correct all winter, and the first days of the series still show 09:00. Colleagues report the mirror problem in the other direction after the October change. Nothing about the meetings themselves was edited. Exp...	--- a/src/schedule.py +++ b/src/schedule.py @@ -12,16 +12,19 @@ def occurrences(meeting, start_date, days): the meeting's own timezone, whatever that instant is in UTC. """ meeting_zone = zone(meeting.tz_name) - first = datetime( - start_date.year, - start_date.month, - start_date...	diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py new file mode 100644 --- /dev/null +++ b/tests/hidden/test_hidden.py @@ -0,0 +1,49 @@ +"""Hidden evaluation tests. Never agent-visible (DEC-015). + +FAIL_TO_PASS: occurrences() steps the series by fixed 24-hour UTC +increments from the first instant...	[ "tests/hidden/test_hidden.py::test_spring_transition_keeps_wall_clock", "tests/hidden/test_hidden.py::test_autumn_transition_keeps_wall_clock", "tests/hidden/test_hidden.py::test_long_series_across_spring_stays_stable" ]	[ "tests/public/test_public.py::test_winter_occurrences_utc_and_local", "tests/public/test_public.py::test_summer_occurrences_hold_local_time", "tests/public/test_public.py::test_agenda_formatting", "tests/public/test_public.py::test_next_occurrence_moves_to_the_following_day", "tests/public/test_public.py::t...	[ { "content": "# Present so pytest adds the repository root to sys.path, making `import src`\n# resolve from the test suite.\n", "path": "conftest.py" }, { "content": "python-dateutil==2.9.0.post0\nsix==1.17.0\n", "path": "requirements.lock" }, { "content": "", "path": "src/__init__.p...	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	250,000	3	6	[ "cheat.diff" ]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_journal_gap_006-v1	pdl_swe_journal_gap_006	1	T3	sequenced journal with contiguity-contract reader	Once the importer rejects a bad row, every row imported afterwards is lost: the writes return normally with no error, but the new entries never show up when the journal is read back — scans and tails behave as if the log ended just before the rejection. A journal that has never rejected anything reads back perfectly. T...	--- a/src/journal.py +++ b/src/journal.py @@ -20,9 +20,9 @@ class Journal: def append(self, record): """Append a record and return its assigned sequence number.""" + payload = encode(record) seq = self._next_seq self._next_seq += 1 - payload = encode(record) self...	diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py new file mode 100644 --- /dev/null +++ b/tests/hidden/test_hidden.py @@ -0,0 +1,44 @@ +"""Hidden evaluation tests. Never agent-visible (DEC-015). + +FAIL_TO_PASS: on the buggy fixture, append() reserves a sequence number +before validation, so a rej...	[ "tests/hidden/test_hidden.py::test_records_after_a_rejection_remain_visible", "tests/hidden/test_hidden.py::test_rejected_records_do_not_consume_sequence_numbers", "tests/hidden/test_hidden.py::test_interleaved_rejections_keep_the_log_complete" ]	[ "tests/public/test_public.py::test_append_and_scan_round_trip", "tests/public/test_public.py::test_append_rejects_invalid_records", "tests/public/test_public.py::test_scan_stops_at_snapshot_gap", "tests/public/test_public.py::test_codec_round_trip_and_validation", "tests/public/test_public.py::test_tail_ret...	[ { "content": "# Present so pytest adds the repository root to sys.path, making `import src`\n# resolve from the test suite.\n", "path": "conftest.py" }, { "content": "# Pure standard library unless the environment image pre-bakes wheels.\n", "path": "requirements.lock" }, { "content": ""...	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	250,000	3	6	[]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_meter_rewind_016-v2	pdl_swe_meter_rewind_016	2	T5	usage metering and billing engine	Finance and support keep hitting the same wall: we cannot say what billing believed at a given moment in the past. When a customer disputed an invoice last month, reconstructing their balance and credit position as of the moment their service was suspended took a day of ad-hoc journal spelunking, the numbers were hand-...	--- a/src/rewind.py +++ b/src/rewind.py @@ -1,17 +1,39 @@ -"""Point-in-time reads over persisted artifacts (not implemented). +"""Point-in-time reads over persisted artifacts. -``rewind(config, store, as_of)`` is intended to answer the question -"what did billing look like at moment X?" from the artifact store -alone...	diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py new file mode 100644 --- /dev/null +++ b/tests/hidden/test_hidden.py @@ -0,0 +1,577 @@ +"""Hidden acceptance suite for point-in-time rewind (never agent-visible). + +Grades rewind() on exact equivalence to the live engine across the +billing query s...	[ "tests/hidden/test_hidden.py::test_tail_replay_of_plain_usage", "tests/hidden/test_hidden.py::test_as_of_at_the_snapshot_moment", "tests/hidden/test_hidden.py::test_simple_close_and_grant_in_tail", "tests/hidden/test_hidden.py::test_accrued_stays_exact_beyond_display_precision", "tests/hidden/test_hidden.py...	[ "tests/public/test_admission_journal.py::test_lateness_window", "tests/public/test_admission_journal.py::test_journal_records_only_admitted_records", "tests/public/test_admission_journal.py::test_segment_rotation_and_names", "tests/public/test_admission_journal.py::test_arrival_clock_is_monotonic_along_the_jo...	[ { "content": "# Present so pytest adds the repository root to sys.path, making `import src`\n# resolve from the test suite.\n", "path": "conftest.py" }, { "content": "# Pure standard library unless the environment image pre-bakes wheels.\n", "path": "requirements.lock" }, { "content": "\...	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml", "src/snapshot.py" ]	500,000	25	38	[ "cheat_class_trust.diff", "cheat_export_reader.diff", "cheat_hand_replay.diff" ]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_plan_drift_013-v2	pdl_swe_plan_drift_013	2	T4	dataflow query planner/optimizer	"Since we moved reporting onto the optimizing planner, some pipelines return wrong results: filters (...TRUNCATED)	"new file mode 100644\n--- /dev/null\n+++ b/src/planner/analysis.py\n@@ -0,0 +1,50 @@\n+\"\"\"Field-(...TRUNCATED)	"diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py\nnew file mode 100644\n--- /(...TRUNCATED)	["tests/hidden/test_hidden.py::test_filter_after_overwriting_derive_uses_final_values","tests/hidden(...TRUNCATED)	["tests/public/test_expr.py::test_arithmetic_and_comparison","tests/public/test_expr.py::test_boolea(...TRUNCATED)	[{"content":"# Present so pytest adds the repository root to sys.path, making `import src`\n# resolv(...TRUNCATED)	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	500,000	10	26	[ "cheat_disable.diff", "cheat_partial.diff", "cheat_typelist.diff" ]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_retry_cap_001-v1	pdl_swe_retry_cap_001	1	T1	HTTP-client retry helper	"Fix the retry function so that the maximum retry count is respected when the wrapped operation rais(...TRUNCATED)	"--- a/src/retry.py\n+++ b/src/retry.py\n@@ -14,7 +14,7 @@\n \"\"\"\n last_exc = None\n (...TRUNCATED)	"diff --git a/tests/hidden/test_retry_hidden.py b/tests/hidden/test_retry_hidden.py\nnew file mode 1(...TRUNCATED)	[ "tests/hidden/test_retry_hidden.py::test_respects_max_attempts_on_timeout" ]	["tests/public/test_retry.py::test_returns_first_success_without_retrying","tests/public/test_retry.(...TRUNCATED)	[{"content":"# Present so pytest adds the repository root to sys.path, making `import src`\n# resolv(...TRUNCATED)	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	250,000	1	2	[]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_retry_thread_008-v1	pdl_swe_retry_thread_008	1	T3	retry orchestration across a dispatch facade	"The attempts option on execute() does nothing. An operation that fails once with a transient error (...TRUNCATED)	"--- a/src/api.py\n+++ b/src/api.py\n@@ -12,4 +12,4 @@ def execute(name, op, attempts=1):\n \"\"(...TRUNCATED)	"diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py\nnew file mode 100644\n--- /(...TRUNCATED)	["tests/hidden/test_hidden.py::test_attempts_retry_until_success","tests/hidden/test_hidden.py::test(...TRUNCATED)	["tests/public/test_public.py::test_execute_success_first_try","tests/public/test_public.py::test_ex(...TRUNCATED)	[{"content":"# Present so pytest adds the repository root to sys.path, making `import src`\n# resolv(...TRUNCATED)	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	250,000	3	6	[]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_route_order_009-v1	pdl_swe_route_order_009	1	T3	web-service request router (route specificity)	"After we reorganized how routes are registered, some API endpoints return the wrong resource: reque(...TRUNCATED)	"--- a/src/router.py\n+++ b/src/router.py\n@@ -44,10 +44,12 @@ class Router:\n def resolve(self,(...TRUNCATED)	"diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py\nnew file mode 100644\n--- /(...TRUNCATED)	["tests/hidden/test_hidden.py::test_daily_report_route_dispatches_to_daily","tests/hidden/test_hidde(...TRUNCATED)	["tests/public/test_public.py::test_users_listing","tests/public/test_public.py::test_users_export_r(...TRUNCATED)	[{"content":"# Present so pytest adds the repository root to sys.path, making `import src`\n# resolv(...TRUNCATED)	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	250,000	3	7	[]	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health
pdl_swe_rule_cascade_017-v2	pdl_swe_rule_cascade_017	2	T5	forward-chaining business-rules engine	"Support escalated two automation incidents from the same customer, both involving rule cascades —(...TRUNCATED)	"--- a/src/agenda.py\n+++ b/src/agenda.py\n@@ -16,7 +16,7 @@ from dataclasses import dataclass\n fro(...TRUNCATED)	"diff --git a/tests/hidden/test_hidden.py b/tests/hidden/test_hidden.py\nnew file mode 100644\n--- /(...TRUNCATED)	["tests/hidden/test_hidden.py::test_firing_uses_values_current_at_fire_time","tests/hidden/test_hidd(...TRUNCATED)	["tests/public/test_cascades.py::test_asserted_fact_triggers_downstream_rule_in_same_drain","tests/p(...TRUNCATED)	[{"content":"# Present so pytest adds the repository root to sys.path, making `import src`\n# resolv(...TRUNCATED)	[ "src/**" ]	[ "tests/**", "requirements.lock", "pyproject.toml" ]	500,000	20	31	["cheat_popcheck.diff","cheat_rebuild.diff","cheat_version_seam_only.diff","cheat_withdraw_only.diff(...TRUNCATED)	0.80hidden_pass_fraction + 0.15public_regression + 0.05*build_health

End of preview. Expand in Data Studio

PDL-SWE-Bench

PDL-SWE-Bench is an agentic software-engineering benchmark maintained by Poindexter Labs — the SWE sibling of PDL-Bench. Each task drops an agent into an original, internally-authored code repository with an engineering issue written as prose, a passing public test suite, and a fixed token budget. The agent's submitted patch is graded against a held-out acceptance suite it never saw during the episode.

Like PDL-Bench, this is an open benchmark (HLE-style): the complete answer key ships with the dataset — the reference (gold) patch, the held-out tests, and the test-level grading contract — so anyone can grade patches locally and reproduce results. See Contamination.

The row schema follows the SWE-bench conventions (problem_statement, patch, test_patch, FAIL_TO_PASS, PASS_TO_PASS) so existing tooling maps over directly, with one difference: these are self-contained synthetic repositories, shipped inline in the files column and as browsable trees under tasks/, rather than references to public GitHub commits. Nothing in this dataset derives from public repositories or issue trackers.

Difficulty tiers

18 tasks span five calibrated tiers. Tier placement is gated by measurement at authoring time, not intuition:

Tier	Tasks	Bar
T1	1	single-file, local reasoning
T2	2	multi-file, requires tracing data flow
T3	10	subtle interactions; typically requires constructing a reproduction
T4	2	complete fix requires building new machinery across files; certified at authoring (2026-08): GPT-5.4-mini and Claude Haiku 4.5 both scored below 1.0
T5	3	flagship tier; certified at authoring (2026-08): Claude Opus 5, GPT-5.6, and Gemini 3.1 Pro all scored below 1.0 under a 500K-token contract while the reference patch scores 1.0

Certification claims are dated, pre-publication facts: they were measured before this dataset was released. Results obtained on these tasks by models trained after publication carry no such evidentiary weight.

Reward model

Episodes are graded on a continuous scale rather than binary resolution:

reward = 0.80 * hidden_pass_fraction + 0.15 * public_regression + 0.05 * build_health

1.0 — complete fix; 0.2 — the "do-nothing floor" (public tests still pass, build healthy); between — genuine partial fixes; below 0.2 — the patch broke working code.
Hidden suites are built as graded ladders: each task shipped with author-replayed partial solutions at known reward values, registered before any model ran (graded_ladder_rungs lists them). In measured campaigns, frontier models repeatedly landed exactly on these pre-registered values — a landed rung names the specific machinery a model failed to build.

Schema

Column	Meaning
`instance_id`, `task_id`, `task_version`	task identity
`tier`, `domain`	difficulty tier and codebase domain
`problem_statement`	the issue text the agent receives
`files`	the full starting repository, `list<{path, content}>`
`patch`	the reference (gold) solution diff
`test_patch`	additive diff placing the held-out tests at `tests/hidden/`
`FAIL_TO_PASS`	held-out test IDs: fail at base, pass with `patch`
`PASS_TO_PASS`	public test IDs: pass at base, must keep passing
`patch_policy_allowed` / `patch_policy_forbidden`	paths the agent may / must not modify (the final workspace state is policed, not just the diff)
`token_budget`	the episode's total token contract
`hidden_test_count`, `public_test_count`	suite sizes
`graded_ladder_rungs`	names of the authored partial-solution rungs
`reward_formula`	the grading formula above

Browsable copies live under tasks/<task_id>/: the task manifest.yaml, the fixture tree, the held-out tests (hidden/), and the reference patch (solution/gold.diff).

Evaluating a model

Materialize files into a working directory (or copy tasks/<task_id>/fixtures/...).
Give the agent problem_statement, the repository, and tools; enforce patch_policy_* and token_budget.
Apply the agent's patch to a pristine copy; apply test_patch; run PASS_TO_PASS and FAIL_TO_PASS (pytest); compute the reward formula.

The numbers published by Poindexter Labs additionally run inside a network-isolated, container-sandboxed harness with full episode provenance (model route, prompt version, image digests, trajectories) and calibrated baselines executed in-campaign (reference patch 1.0, do-nothing agent 0.2, on every task). Evaluation-as-a-service against the same harness — including private, uncontaminated task sets — is available; contact Poindexter Labs.

Contamination

Publishing tasks and answer keys means future training corpora will likely absorb them; that trade-off is accepted deliberately, as with PDL-Bench. These tasks were contamination-clean at authoring (original codebases, original defects, nothing derived from public sources), and all published certification results predate this release. Poindexter Labs maintains unpublished task sets under the same methodology for measurement that must remain contamination-free.

License

CC-BY-4.0.

Downloads last month: 35

Number of rows:

Total file size:

786 kB