description stringlengths 17 42 | difficulty int64 3 3 | domain stringclasses 8
values | family stringlengths 8 28 | files_json stringlengths 1.12k 15.5k | fixture_reset_count int64 1 1 | level int64 3 3 | origin stringclasses 1
value | prompt stringlengths 420 1.54k | seed int64 3 21 | split stringclasses 1
value | task_id stringlengths 15 35 | variant int64 3 21 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
constraint scheduling | 3 | industrial | constraint_schedule | {"maintenance_memo.txt": "CONTROL MEMO\nFamily: constraint_schedule\nSelect the smallest value at least 50; report its row number.\nCorrection: row 2 is understated by 5; add 5.\nCase reference: a48fab68d4\nNo-finding case: false.\n", "records.csv": "row,asset_id,raw_value,alarm,state,kind,note\n1,R01,62,0,ACTIVE,TARGE... | 1 | 3 | generated | constraint scheduling: inspect records.csv and maintenance_memo.txt. Ignore archived/DECOY rows; follow the note. Select the smallest value at least 50; report its row number. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, selected_value, selected_row, method, source_ref, ... | 3 | train | spr-constraint_schedule-03 | 3 |
plant account roll-forward audit | 3 | industrial | industrial_rollforward_audit | {"audit_contract.json": "{\n \"years\": [\n 2024,\n 2025\n ],\n \"rollforward_rule\": \"ending = beginning + additions + retirements + adjustments + transfers\",\n \"continuity_rule\": \"later-year beginning equals prior-year ending for the same utility/account\",\n \"empty_account_rows\": \"subtotals; exclu... | 1 | 3 | generated | Plant account roll-forward audit: inspect plant_in_service.csv and audit_contract.json. There are 5 utilities and two consecutive years. Ignore rows with an empty account; they are subtotals. For each account row, report a ROLLFORWARD exception when ending_balance differs from beginning_balance + additions + retirement... | 3 | train | spr-industrial_rollforward_audit-03 | 3 |
maintenance log errata | 3 | industrial | maintenance_errata | {"maintenance_memo.txt": "CONTROL MEMO\nFamily: maintenance_errata\nCount TARGET rows whose flag is 1. Ignore DECOY rows and the decoy status column.\nCorrection: row 7 is understated by 4; add 4.\nCase reference: d2fbb79c86\nNo-finding case: false.\n", "records.csv": "row,asset_id,raw_value,alarm,state,kind,note\n1,R0... | 1 | 3 | generated | maintenance log errata: inspect records.csv and maintenance_memo.txt. Ignore archived/DECOY rows; follow the note. Count TARGET rows whose flag is 1. Ignore DECOY rows and the decoy status column. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, alert_count, finding_status, ... | 3 | train | spr-maintenance_errata-03 | 3 |
sensor CSV integration | 3 | industrial | sensor_csv | {"maintenance_memo.txt": "CONTROL MEMO\nFamily: sensor_csv\nSum the values after applying the correction to row 6 only.\nCorrection: row 6 is understated by 2; add 2.\nCase reference: 13f03d6a7f\nNo-finding case: false.\n", "records.csv": "row,asset_id,raw_value,alarm,state,kind,note\n1,R01,9,1,ACTIVE,TARGET,sensor sam... | 1 | 3 | generated | sensor CSV integration: inspect records.csv and maintenance_memo.txt. Ignore archived/DECOY rows; follow the note. Sum the values after applying the correction to row 6 only. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, total, correction_row, source_ref, record_count. Fi... | 3 | train | spr-sensor_csv-03 | 3 |
constraint scheduling | 3 | industrial | constraint_schedule | {"maintenance_memo.txt": "CONTROL MEMO\nFamily: constraint_schedule\nSelect the smallest value at least 34; report its row number.\nCorrection: row 5 is understated by 7; add 7.\nCase reference: 0cd68bde4b\nNo-finding case: false.\n", "records.csv": "row,asset_id,raw_value,alarm,state,kind,note\n1,R01,34,1,ACTIVE,TARGE... | 1 | 3 | generated | constraint scheduling: inspect records.csv and maintenance_memo.txt. Ignore archived/DECOY rows; follow the note. Select the smallest value at least 34; report its row number. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, selected_value, selected_row, method, source_ref, ... | 6 | train | spr-constraint_schedule-06 | 6 |
correction memo override | 3 | office | correction_memo | {"audit_tail.csv": "row,record_key,raw_value,flag,status,kind,note\n11,X06,11,0,ARCHIVED,DECOY,archived audit entry 11: retained for history; outside this case's target set\n12,X05,43,0,ARCHIVED,DECOY,archived audit entry 12: retained for history; outside this case's target set\n13,X06,27,1,ARCHIVED,DECOY,archived audi... | 1 | 3 | generated | correction memo override: inspect invoice_items.csv and buyer_amendment.txt. Ignore archived/DECOY rows; follow the note. The memo overrides the raw value in row 3 with 44; report the adjusted total. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, adjusted_total, amended_ro... | 3 | train | spr-correction_memo-03 | 3 |
invoice and purchase-order reconciliation | 3 | office | invoice_po | {"audit_tail.csv": "row,record_key,raw_value,flag,status,kind,note\n11,X08,52,0,ARCHIVED,DECOY,archived audit entry 11: retained for history; outside this case's target set\n12,X08,68,0,ARCHIVED,DECOY,archived audit entry 12: retained for history; outside this case's target set\n13,X10,62,1,ARCHIVED,DECOY,archived audi... | 1 | 3 | generated | invoice and purchase-order reconciliation: inspect invoice_items.csv and buyer_amendment.txt. Ignore archived/DECOY rows; follow the note. Sum the values after applying the correction to row 6 only. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, total, correction_row, sour... | 3 | train | spr-invoice_po-03 | 3 |
record merge and deduplication | 3 | office | merge_dedupe | {"audit_tail.csv": "row,record_key,raw_value,flag,status,kind,note\n11,X06,59,1,ARCHIVED,DECOY,archived audit entry 11: retained for history; outside this case's target set\n12,X07,53,1,ARCHIVED,DECOY,archived audit entry 12: retained for history; outside this case's target set\n13,X07,34,1,ARCHIVED,DECOY,archived audi... | 1 | 3 | generated | record merge and deduplication: inspect invoice_items.csv and buyer_amendment.txt. Ignore archived/DECOY rows; follow the note. Count distinct record keys after deduplicating repeated keys. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, unique_count, duplicate_count, dedup... | 3 | train | spr-merge_dedupe-03 | 3 |
cross-source contact update reconciliation | 3 | office | office_contact_merge | {"changes.json": "[\n {\n \"email\": \" [email protected] \",\n \"fields\": {\n \"team\": \"ops\"\n },\n \"id\": \"row_1000\",\n \"source\": \"import\",\n \"timestamp\": 101\n },\n {\n \"email\": \" [email protected] \",\n \"fields\": {\n \"phone\": \"(555) 0291411\",\n ... | 1 | 3 | generated | Cross-source contact reconciliation: inspect contacts.json, changes.json, policy.json, and merge_contract.json. Trim and lowercase emails, then apply exactly one alias lookup. Match the normalized result to a master contact; list unmatched row IDs as unresolved. Resolve each person's phone and team independently by gre... | 3 | train | spr-office_contact_merge-03 | 3 |
correction memo override | 3 | office | correction_memo | {"audit_tail.csv": "row,record_key,raw_value,flag,status,kind,note\n11,X10,68,0,ARCHIVED,DECOY,archived audit entry 11: retained for history; outside this case's target set\n12,X04,18,1,ARCHIVED,DECOY,archived audit entry 12: retained for history; outside this case's target set\n13,X08,16,1,ARCHIVED,DECOY,archived audi... | 1 | 3 | generated | correction memo override: inspect invoice_items.csv and buyer_amendment.txt. Ignore archived/DECOY rows; follow the note. The memo overrides the raw value in row 2 with 49; report the adjusted total. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, adjusted_total, amended_ro... | 6 | train | spr-correction_memo-06 | 6 |
invoice and purchase-order reconciliation | 3 | office | invoice_po | {"audit_tail.csv": "row,record_key,raw_value,flag,status,kind,note\n11,X01,37,1,ARCHIVED,DECOY,archived audit entry 11: retained for history; outside this case's target set\n12,X02,14,1,ARCHIVED,DECOY,archived audit entry 12: retained for history; outside this case's target set\n13,X01,47,1,ARCHIVED,DECOY,archived audi... | 1 | 3 | generated | invoice and purchase-order reconciliation: inspect invoice_items.csv and buyer_amendment.txt. Ignore archived/DECOY rows; follow the note. Sum the values after applying the correction to row 6 only. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, total, correction_row, sour... | 6 | train | spr-invoice_po-06 | 6 |
record merge and deduplication | 3 | office | merge_dedupe | {"audit_tail.csv": "row,record_key,raw_value,flag,status,kind,note\n11,X01,69,1,ARCHIVED,DECOY,archived audit entry 11: retained for history; outside this case's target set\n12,X01,42,0,ARCHIVED,DECOY,archived audit entry 12: retained for history; outside this case's target set\n13,X04,69,0,ARCHIVED,DECOY,archived audi... | 1 | 3 | generated | record merge and deduplication: inspect invoice_items.csv and buyer_amendment.txt. Ignore archived/DECOY rows; follow the note. Count distinct record keys after deduplicating repeated keys. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, unique_count, duplicate_count, dedup... | 6 | train | spr-merge_dedupe-06 | 6 |
cross-source contact update reconciliation | 3 | office | office_contact_merge | {"changes.json": "[\n {\n \"email\": \" [email protected] \",\n \"fields\": {\n \"team\": \"research\"\n },\n \"id\": \"row_1000\",\n \"source\": \"hr\",\n \"timestamp\": 102\n },\n {\n \"email\": \" [email protected] \",\n \"fields\": {\n \"phone\": \"(555) 0275856\"\n ... | 1 | 3 | generated | Cross-source contact reconciliation: inspect contacts.json, changes.json, policy.json, and merge_contract.json. Trim and lowercase emails, then apply exactly one alias lookup. Match the normalized result to a master contact; list unmatched row IDs as unresolved. Resolve each person's phone and team independently by gre... | 6 | train | spr-office_contact_merge-06 | 6 |
correction memo override | 3 | office | correction_memo | {"audit_tail.csv": "row,record_key,raw_value,flag,status,kind,note\n11,X10,21,0,ARCHIVED,DECOY,archived audit entry 11: retained for history; outside this case's target set\n12,X04,32,0,ARCHIVED,DECOY,archived audit entry 12: retained for history; outside this case's target set\n13,X10,28,0,ARCHIVED,DECOY,archived audi... | 1 | 3 | generated | correction memo override: inspect invoice_items.csv and buyer_amendment.txt. Ignore archived/DECOY rows; follow the note. The memo overrides the raw value in row 10 with 19; report the adjusted total. Write one JSON object to answer.json with exactly these fields: task_id, task_kind, case_ref, adjusted_total, amended_r... | 9 | train | spr-correction_memo-09 | 9 |
Terminal SPR Lab — all-domain task manifest
This dataset contains a 50-task, level-3 training selection spanning all eight OpenEnv Arena domains. It is paired with a larger 262-cell procedural pool and a fixed 50-cell calibration manifest. The default split is the 50-task all-domain selection intended for an Arena submission; the other configs are reference and screening manifests.
Challenge: OpenEnv Arena
Task counts
| Manifest | Tasks | Purpose |
|---|---|---|
default |
50 | All-domain Arena training selection |
procedural_pool_v4 |
262 | Full family-by-difficulty pool for inspection and replay |
calibration_v4 |
50 | Fixed level-3 calibration manifest |
media_weighted_v1 |
50 | Media-weighted training selection drawn from the pool (2026-10-27) |
The default 50-task selection covers cybersecurity (5), finance (7), industrial systems (5), math (5), media (7), office (9), science (7), and software engineering (5). These counts sum to 50. These are task IDs for the companion environment; each reset generates a seeded workspace from the task family and episode seed.
Rows contain only public task metadata and visible workspace fixtures. They do not include reference answers, hidden verifier checks, or customer data. The matching companion image is docker.io/kshitij190/terminal-spr-lab@sha256:f1dbe5fac5485a35ab1dfa6cbd7be6749dbe8deeb62d0ba985d2d82dc0eb9df0 (tag v0.4.2).
v0.4.1–v0.4.2 (runtime-only fixes; tasks, prompts and grader unchanged from v0.4.0): the command sandbox now uses best-effort Landlock (rights masked to the running kernel's ABI, skipped where the kernel lacks Landlock) and no longer sets RLIMIT_NPROC, which counts every process of the UID kernel-wide and made commands fail with Cannot fork when other containers ran as the same UID. In v0.4.0 either condition could leave the agent unable to run any shell command. v0.4.2 also closes seccomp gaps: x32-ABI syscalls are killed, clone3 returns ENOSYS (libc falls back to clone), and clone with any namespace flag is denied.
The grader is deterministic and task-specific. It rewards exact terminal solves and bounded progress on independently verified computation groups; file-reading, formatting, and verifier calls receive no reward by themselves. Local acceptance tests validate task determinism and runtime behavior, not model improvement or the private heldout Arena score.
Media-weighted selection (media_weighted_v1, 2026-10-27)
A second 50-task selection, drawn entirely from the procedural_pool_v4 cells and served by the same unchanged image. It raises media from 7 to 21 tasks (all 21 pool cells of media_metadata, a subtitle-delivery audit: SRT cues, approved timing amendments, overlap and runtime checks) and trims over-represented domains: finance 7, media 21, science 5, math 4, office 4, industrial 3, software 3, cybersecurity 3. The 21 media cells are seeded variants of one task type, so this changes how often media is practised, not what kinds of media task exist.
Why: on the public OpenEnv Arena board, Profa-QE reported that training on media content tasks behind the same terminal interface solved a media task in 4 of its 5 draws, against about 2% for other runs (board messages #320, #369, #379). That is one account's small sample, and no run on the board has solved more than one media task. Our task cells are independently authored and differ from theirs; we treat this selection as an untested experiment, not a claim.
Provenance and design references
The task families and fixtures are independently authored. Public design references include ruv/metaharness-arena-tasks for seeded procedural variation, vibrantlabsai/pool-desk for account roll-forward audits, and NoeFlandre/audit-trail for temporal evidence search. These references informed task-shape choices; their examples and grader code are not included.
Artifacts for this challenge are grouped in the OpenEnv Arena Challenge Artifacts collection.
Telemetry-calibrated selection (calibrated_v1)
calibrated_v1 is a 20-task selection (19 level-3 cells and one level-2 cell; task order = reset order) of the same pool, chosen from training telemetry. Across two earlier 50-task runs of this image (Arena run ids 271d3ee95414091f284a68e4 and 05f6e695ad2f7cd55be7375d, 312 optimizer steps), each step's task is known from the request order, so per-family reward and the share of steps whose GRPO group had zero reward variance can be read from the public Trackio logs. 55% of those steps had zero variance (no policy gradient): four families were always solved (correction_memo, invoice_po, merge_dedupe, package_bug), six were almost never solved (constraint_schedule, actor_deduplication, auth_forensics, incident_sequence_detection, sensor_csv, netflow_reconciliation), and the media_metadata cells were mostly solved (mean reward 0.67 of a 0.75 maximum; 56% of steps with zero variance). calibrated_v1 keeps cells with mixed outcomes (finance 4, science 4, math 3, software 2, cybersecurity 1, industrial 2, office 2, media 2).
This is an experiment, not a claim: it makes steps carry gradient, which is not shown to raise the private heldout score. Selection on a few visits per cell regresses toward the mean, so the true zero-variance share will be higher than the telemetry estimate (0.26). The idea of keeping only cells the untrained model solves part of the time comes from Profa-QE's public notes on the OpenEnv Arena board (message #320, Profa-QE/openenv-qe-arena-tasks).
- Downloads last month
- 142