Datasets:

open-athena
/

task-trove

path stringlengths 14 88	source stringclasses 7 values	family stringclasses 6 values	template_id stringlengths 12 12	converter stringclasses 7 values	mode stringclasses 7 values	dockerfile_id stringlengths 12 12	tags listlengths 2 17	has_solution bool 2 classes	task_binary unknown	solution_binary unknown
contrastive_solver/data-processing_20260415_002933_001.tar.gz	AweAI-Team__CalibForge	terminal-agent	5f13e941694d	calibforge_native_harbor	native-harbor	edff7418f8b5	[ "contrastive_solver", "csv-processing", "data-processing", "data-reconciliation", "financial-data", "matching-algorithms", "native-harbor", "pandas" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 92, 235, 118, 219, 184, 118, 158, 223, 124, 10, 84, 153, 117, 108, 247, 72, 20, 47, 186, 250, 140, 167, 213, 72, 74, 162, 53, 190, 168, 150, 146, 52, 43, 199, 85, 40, 9, 178, 57, ...	null
contrastive_solver/data-processing_20260415_002933_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	3501d9be8bc2	[ "contrastive_solver", "csv-processing", "data-processing", "data-quality", "deduplication", "gap-detection", "native-harbor", "time-series" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 61, 237, 114, 219, 70, 146, 249, 141, 167, 152, 67, 226, 21, 89, 38, 193, 15, 145, 146, 172, 141, 246, 202, 177, 36, 219, 21, 75, 114, 89, 178, 83, 89, 133, 135, 128, 228, 64, 68,...	null
contrastive_solver/data-processing_20260420_020307_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	bc81b73ac8d6	[ "contrastive_solver", "data-aggregation", "data-cleaning", "data-processing", "interpolation", "native-harbor", "time-series", "timezone-handling" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 107, 115, 219, 54, 182, 253, 204, 95, 129, 101, 55, 107, 41, 145, 168, 183, 31, 218, 168, 119, 156, 216, 105, 189, 183, 73, 60, 177, 211, 110, 154, 122, 24, 138, 132, 44, 174, ...	null
contrastive_solver/data-processing_20260420_020307_004.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	05ce0fc3152f	[ "contrastive_solver", "csv-parsing", "data-cleaning", "data-processing", "date-normalization", "encoding", "etl", "native-harbor" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 92, 75, 143, 219, 88, 118, 246, 154, 191, 226, 134, 221, 61, 82, 101, 72, 138, 162, 72, 73, 165, 65, 117, 170, 92, 182, 187, 221, 109, 187, 28, 151, 187, 61, 158, 182, 35, 95, 145...	null
contrastive_solver/data-processing_20260422_235124_001.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	095af4d544b3	[ "contrastive_solver", "data-processing", "data-recovery", "json", "log-analysis", "native-harbor", "python", "text-processing" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 237, 114, 227, 184, 145, 243, 155, 79, 129, 211, 110, 74, 242, 157, 69, 145, 148, 40, 201, 74, 57, 137, 107, 60, 155, 248, 110, 118, 236, 178, 103, 111, 43, 231, 115, 105, 32, ...	null
contrastive_solver/data-processing_20260422_235124_003.tar.gz	AweAI-Team__CalibForge	terminal-agent	48cce96fec1b	calibforge_native_harbor	native-harbor	3ba09e3c01f2	[ "anomaly-detection", "contrastive_solver", "cross-validation", "data-cleaning", "data-processing", "data-quality", "imputation", "native-harbor", "pandas", "sensor-data", "time-series" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 237, 118, 219, 70, 118, 249, 141, 167, 152, 34, 105, 74, 58, 36, 8, 128, 4, 37, 113, 67, 247, 184, 182, 147, 248, 180, 254, 104, 172, 56, 39, 71, 203, 3, 65, 192, 80, 68, 12...	null
contrastive_solver/data-processing_20260424_090344_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	53cdbf6a974e	[ "contrastive_solver", "csv", "data-matching", "data-processing", "financial", "fuzzy-matching", "native-harbor", "reconciliation" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 92, 91, 119, 219, 56, 146, 238, 103, 253, 10, 140, 166, 207, 180, 157, 214, 133, 87, 93, 60, 235, 222, 163, 72, 114, 162, 109, 89, 210, 72, 114, 210, 57, 89, 31, 153, 146, 40, 155...	null
contrastive_solver/data-processing_20260424_090344_003.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	8435f78ee3e2	[ "contrastive_solver", "data-processing", "deduplication", "encoding", "etl", "fuzzy-matching", "native-harbor", "pandas" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 219, 114, 227, 198, 114, 126, 198, 87, 204, 129, 93, 71, 164, 77, 130, 23, 145, 146, 86, 199, 114, 86, 171, 221, 181, 215, 217, 91, 73, 235, 91, 201, 12, 5, 130, 67, 17, 43, ...	null
contrastive_solver/data-processing_20260425_072438_001.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	edd53631f4a5	[ "contrastive_solver", "csv-parsing", "data-cleaning", "data-processing", "encoding", "etl", "native-harbor", "pandas", "schema-normalization" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 91, 91, 115, 219, 56, 150, 238, 103, 253, 10, 140, 186, 107, 100, 239, 136, 20, 117, 243, 77, 227, 221, 36, 118, 210, 227, 233, 78, 39, 21, 103, 58, 219, 21, 123, 21, 138, 132, 36...	null
contrastive_solver/data-processing_20260425_072438_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	12ff0e763b3f	[ "contrastive_solver", "csv-processing", "data-processing", "data-reconciliation", "deduplication", "encoding-detection", "etl", "native-harbor" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 219, 114, 27, 71, 118, 126, 158, 175, 232, 64, 246, 2, 144, 113, 39, 41, 81, 220, 133, 83, 52, 73, 89, 44, 139, 162, 150, 164, 162, 245, 202, 44, 168, 49, 211, 0, 198, 28, 2...	null
contrastive_solver/data-processing_20260427_204204_004.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	f8d4c810a209	[ "contrastive_solver", "data-normalization", "data-processing", "datetime-parsing", "log-processing", "native-harbor", "regex", "timezone-handling" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 107, 83, 227, 56, 182, 253, 217, 191, 66, 235, 153, 222, 14, 219, 216, 177, 243, 224, 145, 45, 246, 22, 183, 97, 182, 123, 239, 208, 116, 53, 244, 78, 13, 44, 101, 20, 91, 33,...	null
contrastive_solver/data-processing_20260428_101637_001.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	d3832efad524	[ "contrastive_solver", "csv", "data-aggregation", "data-processing", "error-handling", "log-parsing", "native-harbor", "regex", "security-analysis" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 105, 115, 219, 70, 150, 249, 140, 95, 209, 203, 120, 151, 82, 134, 4, 1, 144, 32, 45, 57, 180, 71, 35, 43, 137, 119, 28, 91, 99, 41, 147, 76, 121, 84, 80, 19, 104, 146, 176,...	null
contrastive_solver/data-processing_20260428_101637_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	c820810b4fd4	[ "conflict-resolution", "contrastive_solver", "csv-parsing", "data-processing", "data-reconciliation", "encoding-detection", "etl", "json-parsing", "native-harbor" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 203, 114, 227, 70, 146, 62, 243, 43, 106, 224, 241, 136, 180, 9, 136, 79, 61, 168, 208, 172, 213, 106, 181, 173, 177, 187, 213, 35, 181, 199, 225, 104, 43, 208, 69, 176, 40, 1...	null
contrastive_solver/data-processing_20260428_101637_004.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	84a624e71d53	[ "contrastive_solver", "data-processing", "data-validation", "error-handling", "json", "log-processing", "native-harbor", "regex", "text-parsing" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 92, 105, 123, 219, 200, 145, 158, 207, 248, 21, 189, 204, 76, 40, 109, 76, 16, 224, 41, 113, 31, 38, 145, 37, 217, 214, 68, 215, 74, 114, 102, 51, 146, 22, 6, 129, 38, 137, 49, ...	null
contrastive_solver/data-processing_20260429_103257_001.tar.gz	AweAI-Team__CalibForge	terminal-agent	5f13e941694d	calibforge_native_harbor	native-harbor	a79d51b4fa6d	[ "contrastive_solver", "csv-processing", "data-processing", "data-quality", "exception-handling", "financial-data", "native-harbor", "pandas", "transaction-reconciliation" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 92, 123, 115, 219, 70, 146, 207, 223, 248, 20, 115, 140, 183, 34, 229, 72, 16, 79, 130, 224, 158, 246, 202, 231, 56, 27, 215, 197, 150, 202, 114, 46, 119, 229, 85, 209, 32, 57, 18...	null
contrastive_solver/data-processing_20260429_103257_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	d036c84b8d99	[ "contrastive_solver", "data-processing", "data-recovery", "encoding", "error-handling", "etl", "jsonl", "native-harbor" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 61, 235, 122, 227, 182, 149, 249, 173, 167, 64, 153, 108, 37, 181, 186, 95, 108, 143, 186, 206, 238, 100, 236, 73, 102, 119, 102, 236, 207, 246, 36, 219, 122, 92, 154, 18, 33, 155, ...	null
contrastive_solver/data-processing_20260430_081156_004.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	2eee96d8bc0e	[ "character-sets", "contrastive_solver", "csv", "data-processing", "encoding", "file-processing", "native-harbor", "python" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 58, 77, 111, 36, 199, 117, 123, 238, 95, 241, 50, 18, 204, 161, 221, 51, 211, 61, 156, 15, 114, 12, 58, 225, 238, 106, 87, 148, 196, 37, 67, 114, 189, 136, 118, 23, 179, 53, 221, ...	null
contrastive_solver/data-processing_20260501_000112_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	6572209963b1	[ "contrastive_solver", "csv-parsing", "data-processing", "data-reconciliation", "encoding", "etl", "fuzzy-matching", "native-harbor" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 107, 119, 219, 54, 150, 253, 204, 95, 129, 97, 230, 140, 165, 30, 137, 18, 41, 201, 15, 205, 122, 102, 21, 219, 233, 250, 76, 98, 123, 109, 167, 157, 172, 227, 195, 80, 18, 10...	null
contrastive_solver/data-processing_20260501_051538_004.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	3eb27688cbe2	[ "contrastive_solver", "csv", "data-processing", "data-transformation", "debugging", "etl", "json", "native-harbor", "pandas", "python" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 107, 115, 219, 56, 146, 243, 153, 191, 2, 199, 76, 173, 228, 91, 139, 34, 41, 81, 178, 188, 165, 221, 203, 36, 158, 217, 236, 237, 36, 41, 71, 217, 185, 59, 159, 139, 67, 137,...	null
contrastive_solver/data-processing_20260502_090008_001.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	bbe6a28fd05b	[ "algorithm-implementation", "contrastive_solver", "csv-parsing", "data-processing", "data-reconciliation", "etl", "fuzzy-matching", "native-harbor", "unicode" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 107, 119, 219, 54, 150, 253, 172, 95, 129, 85, 187, 35, 105, 135, 164, 72, 234, 237, 25, 207, 172, 106, 59, 137, 103, 226, 71, 45, 167, 221, 158, 140, 143, 66, 145, 144, 197, ...	null
contrastive_solver/data-processing_20260502_090008_002.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	f45750122509	[ "aggregation", "contrastive_solver", "data-processing", "data-quality", "native-harbor", "sensor-data", "time-series", "validation" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 107, 115, 219, 70, 146, 249, 140, 95, 49, 7, 175, 118, 193, 11, 9, 2, 36, 65, 82, 74, 148, 59, 173, 165, 196, 170, 139, 100, 151, 37, 39, 151, 210, 169, 32, 144, 28, 138, 17...	null
contrastive_solver/data-processing_20260502_090008_003.tar.gz	AweAI-Team__CalibForge	terminal-agent	52c23dc22390	calibforge_native_harbor	native-harbor	a94bc63f194b	[ "contrastive_solver", "data-aggregation", "data-processing", "encoding", "etl", "log-parsing", "native-harbor", "pandas", "parquet", "timezone-normalization" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 60, 107, 115, 219, 56, 146, 243, 153, 191, 2, 171, 153, 187, 200, 181, 22, 69, 82, 111, 237, 120, 182, 50, 177, 51, 201, 110, 18, 251, 108, 167, 166, 182, 124, 46, 25, 34, 33, 139...	null
contrastive_solver/data-processing_20260502_090008_004.tar.gz	AweAI-Team__CalibForge	terminal-agent	142b43a830e3	calibforge_native_harbor	native-harbor	90a24ecf9eaa	[ "contrastive_solver", "csv-parsing", "data-processing", "data-validation", "encoding", "etl", "native-harbor", "text-processing" ]	false	[ 31, 139, 8, 0, 0, 0, 0, 0, 2, 19, 237, 91, 79, 115, 219, 72, 118, 159, 235, 162, 42, 223, 161, 23, 51, 181, 36, 19, 0, 4, 40, 82, 146, 181, 165, 137, 100, 201, 158, 245, 236, 200, 118, 73, 118, 166, 166, 108, 133, 110,...	null

End of preview. Expand in Data Studio

TaskTrove Clean

2026-10-08 SweSmith pytest startup update

All 12,720 tasks in laion__swesmith-oracle-filtered-v2 now install the verifyit grading package at Marin commit a9adee624d703250556f8078b3d5beb18b4b57ab. This supersedes the earlier SweSmith verifier update below. Candidate syntax errors now score as failed answers when pytest configuration or trusted tests import broken candidate code before test collection. Missing dependencies and syntax errors in the protected tests themselves remain infrastructure errors. The fix is in Marin PR #9923.

The verifier install pin and configuration change in these task archives. The configuration binds the grader to the existing manifest of files restored by task setup, so it can distinguish protected tests from editable candidate files. Instructions, protected tests, reference solutions, the OAuthlib dependencies added earlier today, and all 930,918 TaskTrove tasks are retained. Other dataset sources retain their immutable source-data revisions and packaged tasks; the ten other Parquet data files are unchanged. The SweSmith entry in manifest.json records the new immutable source-data revision.

The RL Data Atlas requires a quality check using the new packaged grader before clearing its verifier warning. Sources rated Good by the Atlas also require difficulty runs. Previous quality and difficulty reports apply to the earlier dataset revision and cannot validate this update.

2026-10-08 SweSmith verifier update

All 12,720 tasks in laion__swesmith-oracle-filtered-v2 now install the verifyit grading package at Marin commit 302649433c75c7199372a26b0346445b6d08f437. Candidate syntax errors now score as failed answers. Missing dependencies and errors collecting protected test code remain infrastructure errors. The fix is in Marin PR #9923.

Only the verifier install pin changes in these task archives. Instructions, protected tests, reference solutions, the OAuthlib dependencies added earlier today, and all 930,918 TaskTrove tasks are retained. Other dataset sources retain their immutable source-data revisions and packaged tasks; the ten other Parquet data files are unchanged. The SweSmith entry in manifest.json records the new immutable source-data revision.

The RL Data Atlas requires a quality check using the new packaged grader before clearing its verifier warning. Sources rated Good also require difficulty runs. Previous quality and difficulty reports apply to the earlier dataset revision and cannot validate this update.

2026-10-08 SweSmith OAuthlib dependencies

This release adds the missing PyJWT, cryptography, and blinker test dependencies to the packaged environments of 66 OAuthlib tasks in laion__swesmith-oracle-filtered-v2. It retains all 12,720 SweSmith tasks and all 930,918 TaskTrove tasks.

For the reported task swesmith-00022, the original environment could not collect the protected tests even with correct reference code. With the dependencies installed, native Docker grading passes all 673 selected tests for that task's reference code and returns a score of zero for its mutated code. Grading ran with networking disabled.

The repair is in Marin PR #9870. The SweSmith entry in manifest.json records its immutable source-data revision. Other source pins and task counts, the ten other Parquet data files, and other repository files are preserved.

The RL Data Atlas requires fresh QC for the new SweSmith revision before clearing its verifier hold and publishing applicable difficulty results. That hold applies to Atlas QC and difficulty reports; it does not restrict access to the dataset files.

2026-10-07 SweSmith project setup follow-up

This release updates the environments of all 12,720 laion__swesmith-oracle-filtered-v2 tasks and retains all 930,918 TaskTrove tasks.

The shared pip constraint keeps pytest below 9 while allowing each project's pinned version. This fixes the conflict with DeepDiff's pytest==8.3.4 development requirement. Conan task images also install the missing test dependencies: mock, webtest, PyJWT, bottle, and parameterized. The earlier tldextract plugin and marshmallow simplejson repairs remain in place. Repository pytest plugins stay enabled.

The code is in Marin PR #9870. The SweSmith entry in manifest.json records the immutable source-data revision. All other source pins and task counts, the ten additional-source data files, and the other repository files are preserved.

The RL Data Atlas requires fresh SweSmith QC on this revision before its verifier hold can be resolved or difficulty checked. This release does not change the applicability of other sources' current reviews or difficulty reports.

2026-10-07 Python contract follow-up

This release updates two existing task archives and retains all 930,918 TaskTrove tasks.

DCAgent__exp_rpt_e2egit-v2/e2egit-0289: inventory grading checks the names, quantities, and prices returned after updates. The instruction specifies that an Item.decrease_quantity operation that would reduce stock below zero preserves stock and may return unchanged or raise ValueError; invalid amounts in Item.increase_quantity and Inventory.remove_item raise ValueError or TypeError. All 487 E2EGit tasks are retained.
DCAgent__exp_rpt_stack-pytest-v2/stack-pytest-0249: the AlignGraph task defines its row-vector rotation convention and public cliquematch.wrappers.aligngraph.MaskFilter interface without disclosing fixed test answers or an implementation algorithm. Grading retains the three original cases, uses a supported NumPy boolean dtype, and checks other rotations and translations. NumPy is installed for both grading and solver Python commands. All 495 Stack tasks are retained.

The repairs are in Marin PR #9869 and Marin PR #9871. Their entries in manifest.json record immutable source-data revisions. All other source pins and task counts are preserved, including the NL2Bash follow-up and ten additional-source data files.

The RL Data Atlas quality assessments for these two sources have not yet been refreshed for this revision. Their source-level difficulty ratings remain withheld pending fresh QC.

2026-10-07 NL2Bash contract follow-up

This release updates two existing tasks in DCAgent2__nl2bash-tasks-cleaned-oracle-v2 and retains all 1,497 tasks in that source and all 930,918 tasks in TaskTrove.

task_1492 names /workspace/ruby_samples.txt as the first file and /workspace/sort_samples.txt as the second. Grading requires lines unique to the second file in descending numeric order; the reversed difference no longer earns credit.
task_3002 requires the created text file to be nonempty and contain at least one lowercase o. Grading checks its character dump after replacing lowercase o with 0, and rejects empty captures and unchanged input. Alternative input filenames and text, and dump formats with or without offsets, remain supported.

Marin PR #9866 contains the repairs. manifest.json records the source's immutable data revision. All other source-data pins and task counts are preserved, including the E2EGit repairs and ten additional-source data files.

NL2Bash's source-level quality assessment in the RL Data Atlas must be refreshed for this revision. Difficulty remains withheld pending fresh QC.

2026-10-07 E2EGit contract follow-up

This release repairs the packaged instructions and grading files for three existing tasks in DCAgent__exp_rpt_e2egit-v2. It retains all 487 tasks in that source and all 930,918 tasks in TaskTrove.

e2egit-0289 (inventory): Item.decrease_quantity may reject a reduction below zero with ValueError if the quantity stays unchanged. Grading also checks that increasing or removing quantities rejects zero, negative, and non-integer amounts.
e2egit-0359 (calculator): instructions specify calculator.py; grading checks addition, subtraction, multiplication, division, floating-point inputs and outputs, and method documentation.
e2egit-0328 (factorial): student tests belong in /app/tests/test_factorial.py, away from private verifier files. Grading checks recursion, documentation, type hints, and that the student's tests detect errors in negative-input validation, zero, and each integer from one through five. Factorial results are also checked independently.

Marin PR #9869 contains the repairs. The source entry in manifest.json records its immutable data revision. All other source-data pins, task counts, and the ten additional-source data files are preserved. The earlier verifier repair for the JavaScript/TypeScript todo task, e2egit-0433, is retained.

E2EGit's source-level quality assessment in the RL Data Atlas must be refreshed for this revision. Difficulty remains withheld pending fresh QC.

2026-10-07 verifier follow-up

This follow-up changes the packaged grading files for 6,859 existing tasks across three sources and retains all 930,918 tasks.

Source column value	Repair
`DCAgent__code-contests-noblock`	Honor 6,856 stated execution time limits, including fractional seconds. Tasks with unavailable limits retain the prior timeout.
`DCAgent2__nl2bash-tasks-cleaned-oracle-v2`	Repair two oracle mismatches: accept lines unique to the second input file in descending numeric order, and accept character dumps with or without offsets for any text input left in the workspace.
`DCAgent__exp_rpt_stack-pytest-v2`	Enforce the fake Selenium task's stated country data, zone details, and back navigation; empty driver rows no longer earn full credit.

The corresponding entries in manifest.json record their repair PRs and immutable source-data revisions. All other source-data pins, task counts, and ten additional-source shards are preserved. The earlier seven-source repair is retained.

Previous QC ratings and difficulty reports have not been revalidated for these three updated sources. Fresh QC is required for the new data revision.

2026-10-07 verifier repairs

This release repairs 75,775 task archives across seven existing sources while retaining all 930,918 tasks.

Source column value	Repair
`DCAgent2__nl2bash-tasks-cleaned-oracle-v2`	Reject extra line counts that contradict the expected count for the same file or total.
`DCAgent__code-contests-noblock`	Restore comparison after stripping outer whitespace.
`laion__all-puzzles-v2`	Recompute sorting answers and oracle solutions from the prompt's word list and direction.
`DCAgent__exp_rpt_e2egit-v2`	Grade the todo-list task in JavaScript/TypeScript, with separate test processes and shared state within each scenario.
`DCAgent__exp_rpt_stack-pytest-v2`	Supply promised StreamFile test helpers, repair its invalid `pytest.raises` call, and install `mock`.
`laion__swesmith-oracle-filtered-v2`	Use compatible pytest 8 releases and install `simplejson` for marshmallow tasks.
`laion__nemotron-gym-structured-outputs-v4`	Require field values supported by the document. Missing document values must use `__MISSING__`, retaining required field names.

The repaired archives install the verifier from Marin e9859a4e80. Each affected source's entry in manifest.json links its repair PR. All ten shards listed under additional_sources in that manifest, including MiMo, OpenSWE, SWEGym, R2EGym, and CalibForge, are unchanged. Unrelated task archives and their pinned source-data revisions are also unchanged.

Existing ratings for these seven sources apply to earlier data revisions. New quality checks are pending; difficulty will be measured for sources that pass QC.

The current train split contains 930,918 tasks. The original conversion report below describes the original population. Native additions are documented in their source sections; manifest.json contains the combined counts and provenance.

Version note: 2026-10-07

Added the remaining MiMo-V2.6-RL-oss configurations: cyber (1,000 tasks), general (989 tasks) and webdev (2,093 tasks). Task Trove now includes all five MiMo configurations, totaling 7,780 tasks. The train split contains 930,918 tasks across 103 source entries; native additions account for 69,070 tasks across ten source entries beyond the original conversion.

Quality review remains pending for these three additions. Full archive audits check every task against its original inputs and pinned native packages; selected runtime controls are documented below. Inventory inclusion does not establish a Good rating. General retains 64 terminal_bench tasks that the pinned upstream runtime cannot execute; its other 925 tasks use MCP environments. See each configuration's provenance and control limits in the source sections.

The public RL Data Atlas has current difficulty comparisons for all 42 sources rated Good by its quality reviews. The Eurus Small-model comparison still retains one unverified attempt after native sandbox state loss; that attempt receives no score, as detailed in the October 6 note.

Version note: 2026-10-06

Native additions contain 64,988 added tasks across seven source entries, beyond the 861,848-task original conversion documented below. The train split now contains 926,836 tasks across 100 source entries; the two OpenSWE configurations count as separate entries.

Added source / configuration	Tasks
CalibForge	5,431
SWE-Gym	2,438
R2E-Gym-V1	8,101
MiMo-V2.6-RL-oss — music	1,000
MiMo-V2.6-RL-oss — code	2,698
OpenSWE — openswe_oss	36,884
OpenSWE — openswe_other	8,436

The added tasks preserve upstream grading behavior and reference specific upstream revisions. Older task shards are byte-identical. OpenSWE quality review is pending: in one reviewed openswe_other task, the native grader awarded full reward even though a canonical test failed. Inventory inclusion does not imply quality approval; see the source sections below and the RL Data Atlas for current review status.

MiMo's music and code configurations are included. Code quality review is pending: canonical no-op and empty-commit controls on one task matched separate direct native runs with reward 0; synthetic controls validate the adapter's behavior. This does not establish quality across all 2,698 tasks. At this October 6 snapshot, MiMo's cyber, general and webdev configurations were not yet included; native runtime validation and publication were pending.

As of 2026-10-06 at 21:05 UTC, the public RL Data Atlas has current difficulty comparisons for all 42 sources rated Good by its quality reviews. These source entries come from Task Trove and MarinSkyRL. Difficulty comparisons report each model's fraction of sampled tasks solved. Current comparisons use the same sampled tasks per source across Small Qwen3-Coder-30B-A3B-Instruct, Large Qwen3.5-122B-A10B (thinking disabled) and Hosted GLM-5.3 with low reasoning effort, graded by each source's current native grader. Completing and publishing these comparisons was part of the Atlas update's scope. Eurus includes one unverified Small-model attempt: its native sandbox lost state after its generated program reached the six-second execution time limit. The native error and model response are retained; the attempt receives no score. Eurus solve rates use 31 verified Small-model attempts and 32 each for Large and Hosted. Eurus's standard-input line lists were serialized with newlines using the source grader's documented behavior before loading tasks into the native environment; all 32 originally sampled records were retained.

Original Task Trove conversion

TaskTrove Clean is a normalized release of open-thoughts/TaskTrove at revision 0292300. It contains 861,848 retained Harbor tasks from 1,739,326 input rows. The graders were built from Marin commit b76d03131c.

How it was made

The conversion pipeline applies these stages:

Pin the upstream Hugging Face revision and inventory each source's task templates.
Keep sources with recoverable task contracts and record every source decision.
Convert each retained row to the common Harbor layout and replace its source grader with a typed tests/verifier.toml contract executed by tasktrove-verify.
Deduplicate exact instructions within each source.
Reject tasks with a malformed contract, missing verifier files, legacy grader dependencies, exposed solutions or long gold answers, or an invalid mode-specific shape.
For grader modes with safe probes, reject a nonzero score for an empty answer, a non-unit score for the oracle answer, or a positive score for a negative perturbation.
Write retained tasks to the train split and every rejected row to ledger.parquet.

Conversion and cleanup are deterministic. Some published tasks declare an LLM judge for evaluation time. The release report below records the exact converters, grader modes, source decisions, rejection checks, and Docker environments in this build.

Row schema

Each row represents one retained task:

path is the stable identifier from the upstream dataset.
source, family, template_id, converter, mode, dockerfile_id, language, tags, and has_solution are selection and provenance fields.
task_binary is a gzip-compressed Harbor task archive.
solution_binary is an optional, separate gzip-compressed oracle-solution archive.

For the original converted population, the task archive contains instruction.md, task.toml, environment/Dockerfile, tests/test.sh, tests/verifier.toml, and the hidden files needed by its declared grader. The oracle solution is never included inside task_binary.

Loading the tasks

from datasets import load_dataset

tasks = load_dataset("open-athena/task-trove", split="train")
task = tasks[0]
print(task["path"], task["source"], task["mode"])

Filter on the ordinary metadata columns before opening task_binary when selecting a source, grader, language, or tag cohort.

Audit metadata

ledger.parquet contains source, path, status, and error for every rejected input row.
manifest.json contains the pinned upstream revision, verifier commit, aggregate counts, source policy, and Dockerfile inventory.
The original report below describes the original converted population; the manifest includes the native CalibForge addition.

License

The pinned upstream TaskTrove dataset declares the Apache 2.0 license. The source column and release manifest retain source-level provenance for reviewing the terms that apply to a selected cohort.

Release report

861848 of 1739326 tasks from open-thoughts/TaskTrove @ 0292300, graded by tasktrove-verify @ b76d03131c: 43 of 93 sources, 19 converters, 12 modes, 39 distinct Dockerfiles.

By status

status	tasks
converted	861848
routed:sft	458562
dropped_source	212418
routed:garbage	120868
duplicate	54306
unsupported_variant	11929
routed:missing	8408
gold_in_instruction	4850
null_grader	2703
verified:empty	2433
verified:gold_leak	990
reviewed_defect	11

Routed MCQA tasks

route	tasks
sft	458562
garbage	120868
rl	23860

Kept sources

source	family	tasks	clean	main reason for the rest
DCAgent2__nl2bash-tasks-cleaned-oracle-v2	shell-cmd	1498	1497	reviewed_defect (1)
DCAgent__code-contests-noblock	competitive-programming	8728	8222	gold_in_instruction (504)
DCAgent__exp_rpt_curriculum-easy	unit-test-gen	509	505	null_grader (4)
DCAgent__exp_rpt_curriculum-medium-v2	unit-test-gen	492	489	null_grader (3)
DCAgent__exp_rpt_e2egit-large	unit-test-gen	4998	4993	unsupported_variant (5)
DCAgent__exp_rpt_e2egit-v2	unit-test-gen	487	487
DCAgent__exp_rpt_multifile-v3	unit-test-gen	4843	4842	null_grader (1)
DCAgent__exp_rpt_pymethods2test-large-v2	unit-test-gen	4991	4990	reviewed_defect (1)
DCAgent__exp_rpt_pymethods2test-v3	unit-test-gen	500	500
DCAgent__exp_rpt_stack-pytest-v2	unit-test-gen	500	495	unsupported_variant (4)
DCAgent__exp_rpt_unitsyn-python-large-v2	unit-test-gen	4991	4991
DCAgent__exp_rpt_unitsyn-python-v4	unit-test-gen	491	491
DCAgent__swe_rebench_v2_patched_oracle-v2	swe-repo	18319	13559	unsupported_variant (4756)
SankalpKJ__nemotron-math-oracle-filtered-v2	math-answer	57777	57383	verified:empty (364)
laion__all-puzzles-v2	math-answer	6926	6719	verified:gold_leak (202)
laion__codeforces-v3	competitive-programming	10000	9697	null_grader (240)
laion__exp_rpt_taco-v2	stdin-stdout	10000	5116	gold_in_instruction (2592)
laion__glaive-code-assistant-sandboxes-verified-v2	llm-judge-freeform	10000	9994	duplicate (6)
laion__nemo-prism-math-v3	math-answer	2404	2219	duplicate (182)
laion__nemotron-gym-agent-calendar-v2	tool-use	2699	2699
laion__nemotron-gym-arc-agi-python-inductive-v2	other	10000	10000
laion__nemotron-gym-arc-agi-transductive-v3	other	10000	9994	verified:gold_leak (6)
laion__nemotron-gym-competitive-coding-v2	competitive-programming	15713	13973	gold_in_instruction (1739)
laion__nemotron-gym-instruction-following-calendar-v3	instruction-following	5673	5673
laion__nemotron-gym-instruction-following-structured-v3	instruction-following	9437	9167	unsupported_variant (254)
laion__nemotron-gym-instruction-following-v3	instruction-following	46391	46391
laion__nemotron-gym-knowledge-mcqa-v2	qa-short-answer	616888	23860	routed:sft (458562)
laion__nemotron-gym-knowledge-openqa-v4	qa-short-answer	122357	121961	verified:gold_leak (350)
laion__nemotron-gym-math-openmathreasoning-v2	math-answer	42636	42506	verified:empty (102)
laion__nemotron-gym-math-stack-overflow-v3	math-answer	110730	110266	verified:gold_leak (232)
laion__nemotron-gym-math-v5	math-answer	4096	3891	verified:empty (157)
laion__nemotron-gym-multichallenge-advanced-v4	llm-judge-freeform	1068	1052	duplicate (16)
laion__nemotron-gym-reasoning-gym-v2	other	14259	13712	unsupported_variant (282)
laion__nemotron-gym-safety-v3	llm-judge-freeform	89066	44884	duplicate (44182)
laion__nemotron-gym-science-so-openq-v3	llm-judge-freeform	150644	150468	verified:gold_leak (176)
laion__nemotron-gym-structured-outputs-v4	other	53870	50446	null_grader (2438)
laion__stackexchange-codereview-sandboxes-verified-v2	llm-judge-freeform	10000	10000
laion__stackexchange-overflow-sandboxes-verified-v2	llm-judge-freeform	10000	10000
laion__stackexchange-superuser-sandboxes-verified-v2	llm-judge-freeform	10000	10000
laion__stackexchange-tezos-sandboxes-verified-v2	llm-judge-freeform	10000	997	duplicate (9003)
laion__stackexchange-unix-sandboxes-verified-v2	llm-judge-freeform	10000	10000
laion__swesmith-oracle-filtered-v2	swe-repo	12927	12720	unsupported_variant (207)
laion__wizardlm-orca-v4	llm-judge-freeform	10000	9999	reviewed_defect (1)

Dropped sources

source	family	tasks	reason
DCAgent__exp_rle_adversarial-v6	unit-test-gen	2726	The legacy pytest grader performs source-specific Django discovery and loads extra plugins. Ten sampled empty and trivial submissions failed, but the source has no oracle and needs a custom environment adapter.
DCAgent__exp_rpt_nemotron-cpp	unit-test-gen	4196	GoogleTest tasks without shipped oracles. Sampled empty and trivial submissions failed, but some tests contain the reference implementation and others require an uninstalled doctest dependency.
DCAgent__inferredbugs-sandboxes-verifier	swe-repo	9659	Never compiles or runs; regex on the rewritten method body with guards that accept either polarity.
DCAgent__mix_h4_binary_easy	unit-test-gen	1996	Mixed verifier shapes: eight of 10 sampled empty and trivial checks timed out, and the crosscodeeval slice only checks that an import succeeds.
DCAgent__selfinstruct-naive-sandboxes-2-verified-v3	shell-cmd	6665	Per-task LLM-written test_state.py with loose file discovery and dead code. Task ideas are usable; regenerate verifiers with an oracle/no-op gate.
SankalpKJ__nemotron-code-oracle-filtered	competitive-programming	15165	Only test is the example shown in the prompt. Oracle solutions exist, so generate hidden cases by fuzzing inputs through the oracle.
laion__codeelo-v2	competitive-programming	500	Byte-identical generator to codeforces-v3 at 500 rows; merge, do not keep separately.
laion__exp_rpt_bugsinpy-v4	swe-repo	479	LLM-synthesized tests against a single-file stub, with assert True placeholders. Rewrite against the real BugsInPy project suites.
laion__exp_rpt_codenet-python-v4	competitive-programming	6975	Only 3 hidden cases and whitespace-collapsing compare. Oracle present; regenerate 20+ cases per task.
laion__exp_rpt_crosscodeeval-csharp-v4	qa-short-answer	1768	0.25 reward for any identifier-shaped output; instruction coaches the hack.
laion__exp_rpt_crosscodeeval-java-v3	qa-short-answer	2139	Exact string match on a single line completion. Not agentic, no execution.
laion__exp_rpt_crosscodeeval-python-v2	other	500	0.25 for any non-empty output; instruction discloses the tiers.
laion__exp_rpt_crosscodeeval-typescript-v2	other	3356	Re-skin of the Python variant with the same free 0.25 tier; metadata still says python.
laion__exp_rpt_ghactions-v3	other	9930	Instruction lists every job and step verbatim; workflow is never executed. Transcription.
laion__exp_rpt_methods2test-large-v4	unit-test-gen	1194	All 10 sampled shipped Java oracles failed because Maven could not resolve its plugins offline. Recovering the source requires rebuilding the Java grading environment.
laion__exp_rpt_nemotron-junit-v6	unit-test-gen	447	20% of sampled tasks contain unconditional fail() stubs the verifier restores.
laion__exp_rpt_scaffold-v3	other	3121	LLM-synthesized stub-filling toys (TypeScript formatter shim, Flask hello page). Kata-grade.
laion__exp_rpt_stack-cpp-v4	unit-test-gen	7878	Tests are lifted from real repositories with the repository stripped: sampled tasks include headers and data files that do not exist in the image, and one pastes the reference Solution class inside the test.
laion__exp_rpt_stack-dockerfile-gpt5mini-v7	tool-use	587	587 rows of gpt-5-mini-written per-task test scripts whose instructions describe containers that do not exist.
laion__exp_rpt_stack-go-v5	unit-test-gen	2275	Tests import packages from the original repository (bridgr/internal/..., gosnowflake internals) that are not in the task, so most tasks are unsolvable as specified.
laion__exp_rpt_stack-jest-v5	unit-test-gen	424	Spy-call contracts against a 90-package global npm image; 424 rows.
laion__exp_rpt_stack-junit-v6	unit-test-gen	843	JUnit grading is real but every instruction cites a test path that does not exist and scan-class-path counts any test class.
laion__exp_rpt_stack-php-large-v9	unit-test-gen	462	Fail-open exit paths, regex class discovery, 462 rows.
laion__exp_rpt_stack-pytest-large-v3	unit-test-gen	1782	Stripped-repository pytest tasks without shipped oracles. Two of 10 sampled empty and trivial checks timed out, and sampled tests include truthiness-only assertions that weak stubs can satisfy.
laion__exp_rpt_stack-rspec-v4	unit-test-gen	8860	Real Ruby test files but gems are never installed and some tasks are unsolvable offline. Bake gems and drop tasks that fail the oracle gate.
laion__magicoder-v4	llm-judge-freeform	4096	Judge-only over a bundle of every file under /app collected by tests/collect_submission.py; the judge mode grades one answer file, so the bundle shape needs its own converter. Vague refactor prompts, nothing executed.
laion__mix_h10_reward_proportional-v2	unit-test-gen	2858	Four of 10 sampled trivial submissions received full credit because the codereval slice tests a local mock rather than the solution.
laion__mix_h11_single_skill_only-v2	unit-test-gen	2859	Mixed verifier shapes: eight of 10 sampled empty and trivial checks timed out; the source also contains content-free crosscodeeval slices and syntactically invalid tests.
laion__mix_h8_original_tests-v2	unit-test-gen	2848	All 10 sampled empty and trivial checks timed out, and prior inspection found import-only test files in three of 10 tasks.
laion__nemotron-gym-agentic-function-calling-pivot-v3	tool-use	9579	Predict-the-next-call from a frozen transcript with exact key-set match. The only tool-call data here; rewrite into executable mock-tool envs built from the transcripts.
laion__nemotron-gym-agentic-indirect-prompt-injection-v3	prompt-injection	1272	Five of five reviewed rows use the same negative-only checker, which rewards a generic reply or unrelated action without validating the required safe continuation.
laion__nemotron-gym-agentic-swe-pivot-v4	tool-use	1541	No repo in the container; a 9B judge rates one predicted next action.
laion__nemotron-gym-cfbench-v4	instruction-following	468	Deterministic gate before the judge uses 31 constraint ids outside the IFEval registry (tables, heading depth, numbered lists, unique words, ...); only 328 of 1,478 tasks are gate-able today. Port the gate checks before converting.
laion__nemotron-gym-identity-following-v4	instruction-following	21660	Persona is NVIDIA's; judge-only. Rewrite with our identity and deterministic name/language checks, or skip.
laion__nemotron-gym-instruction-following-adversarial-v5	instruction-following	1000	Asks an LLM judge to count exactly five spelling errors.
laion__nemotron-gym-instruction-following-citation-v2	instruction-following	9033	Grades presence of literal marker substrings; never checks the cited content.
laion__nemotron-gym-instruction-following-freeform-v2	instruction-following	8869	Counts markdown tables and bullets; no content check.
laion__nemotron-gym-instruction-following-multiturnchat-v4	instruction-following	1982	Required literal format contradicts the demonstrated turns; judge-only.
laion__nemotron-gym-inverse-ifeval-v4	instruction-following	1000	Gate matches against a deliberately broken synthetic reference; then judge.
laion__nemotron-gym-knowledge-web-search-mcqa-v2	qa-short-answer	2915	Promises web search but ships no tool. Worth rewriting as a real search-tool env; otherwise it is a 3k duplicate of mcqa.
laion__nemotron-gym-litmus-bench-v2	math-answer	5232	Instruction asks for ((answer)), verifier greps boxed or last number; SMILES tasks with no RDKit. Fix format contract and install cheminformatics.
laion__nemotron-gym-math-advanced-calculations-v4	math-answer	5291	Instruction refers to tools that do not exist and only the last number is graded. Ground-truth expression tree is present, so rewrite with a calculator tool and grade every subexpression.
laion__nemotron-gym-multichallenge-vanilla-v3	llm-judge-freeform	1050	Single subjective criterion with 'Expected answer: YES' embedded in the judge prompt.
laion__nemotron-gym-qa-abstention-v4	qa-short-answer	3150	Abstention is never rewarded so the framing is dead, reference leaks into judge text, and it duplicates openqa.
laion__nemotron-gym-sysbench-v4	instruction-following	1010	Deterministic gate before the judge uses 31 constraint ids outside the IFEval registry (tables, heading depth, numbered lists, unique words, ...); only 328 of 1,478 tasks are gate-able today. Port the gate checks before converting.
laion__openswe-tasks-patched-v7-oracle-success	swe-repo	11730	No FAIL_TO_PASS ids: the v7 verifier scores whichever tests its custom pytest guard plugin saw execute, and the repository is cloned by a root-level setup script at agent time. Needs its own converter.
laion__r2egym-patched-full-oracle-v3	swe-repo	2574	Grades by overlap between test_info.json and expected_output_json rather than by pytest node id; not the trusted-paths shape the swe converters handle.
laion__swegym-tasks-patched-validated-v5	swe-repo	2428	Image ships no repository: the instruction clones it and runs make init at agent time, and the old grader pip-installed requirements again at grading time. Sampled oracles fail on missing dependencies and the empty check cannot start.
laion__toolscale-v4	tool-use	4048	Good design (offline tool service) but the CLI script embeds the gold calls and answer, and the prompt states the conclusion. Move the fixture behind a server and strip the success criteria.
laion__tulu3-sft-personas-math-sandboxes-verified-v3	math-answer	9998	Easy SFT persona math, gold in plaintext, carries the terminal-bench canary.

By converter

converter	clean	rejected	dockerfiles
agent_calendar	2699	0	1
all_puzzles	6719	207	2
code_contests	8222	0	1
codeforces	9697	48	1
judge_rubric	105874	53191	4
nemotron_competitive	13973	0	1
nemotron_if_structured	14840	0	2
nemotron_ifeval	46391	0	1
nemotron_math	216265	1377	2
nemotron_mcqa	23860	589420	1
nemotron_multichallenge	1052	16	1
nemotron_openqa	272429	572	1
nemotron_reasoning	33706	271	3
nemotron_structured_outputs	50446	462	2
nl2bash	1497	0	1
python_unit_tests	22783	0	5
swe_patched	13559	3	14
swe_trusted_paths	12720	0	2
taco	5116	0	1

Dockerfiles

id	base image	tasks	converters	sources
e9270f2c1573	python:3.12-slim-bookworm	318365	nemotron_openqa (272429), judge_rubric (44884), nemotron_multichallenge (1052)	4
7ebf98ecd2fd	python:3.11-slim-bookworm	212374	nemotron_math (212374)	4
7aee1955b736	python:3.11-slim-bookworm	102590	nemotron_ifeval (46391), nemotron_mcqa (23860), nemotron_competitive (13973), nemotron_reasoning (9994), nemotron_if_structured (5673), agent_calendar (2699)	6
049de4a4051e	ubuntu:24.04	40997	judge_rubric (40997)	5
49879a65e274	python:3.11-slim-bookworm	33631	nemotron_structured_outputs (33631)	1
5f4521cd7ffa	python:3.11-slim-bookworm	16815	nemotron_structured_outputs (16815)	1
6d77b5c24d4c	python:3.11-slim-bookworm	13712	nemotron_reasoning (13712)	1
6bbc92a19d6e	python:3.10-bookworm	12191	swe_trusted_paths (12191)	1
6c4268eaf683	python:3.10-slim	11467	python_unit_tests (11467)	5
511528c45654	python:3.11-slim-bookworm	10000	nemotron_reasoning (10000)	1
161f75c9e8c0	ubuntu:24.04	9999	judge_rubric (9999)	1
a586757c3643	ubuntu:24.04	9994	judge_rubric (9994)	1
43963417e109	python:3.10-slim	9697	codeforces (9697)	1
58af8699e4d0	python:3.11-slim-bookworm	9167	nemotron_if_structured (9167)	1
fceb5c7f23dc	ubuntu:24.04	8222	code_contests (8222)	1
ae9228630cae	python:3.10-slim	5480	python_unit_tests (5480)	2
5458c53ac683	python:3.11-slim-bookworm	5331	python_unit_tests (5331)	2
3f4ad4c6f411	python:3.10-slim	5116	taco (5116)	1
72298b522224	python:3.10-bookworm	4834	swe_patched (4834)	1
629cc1983f1c	golang:1.23-bookworm	4086	swe_patched (4086)	1
30aacebaf171	python:3.11-slim-bookworm	3891	nemotron_math (3891)	1
59f4a4193b71	python:3.10-slim	3669	all_puzzles (3669)	1
816f4d2c91cd	python:3.10-slim	3050	all_puzzles (3050)	1
b86179e3733c	rust:1.84-bookworm	1837	swe_patched (1837)	1
fd46f9c11b2f	ubuntu:24.04	1497	nl2bash (1497)	1
a1daae1734cb	eclipse-temurin:21-jdk-jammy	1133	swe_patched (1133)	1
b17e642d82f8	python:3.9-bookworm	638	swe_patched (638)	1
63cea345c259	julia:1.10-bookworm	599	swe_patched (599)	1
46d84a733986	python:3.11-bookworm	529	swe_trusted_paths (529)	1
35cb29377a88	ubuntu:24.04	495	python_unit_tests (495)	1
848ec1dbfa78	swift:5.10	266	swe_patched (266)	1
09816786530c	gcc:13-bookworm	50	swe_patched (50)	1
68fff6718dd6	dart:stable	46	swe_patched (46)	1
a6513b93c96e	php:8.3-bookworm	18	swe_patched (18)	1
28f059c03ac4	mcr.microsoft.com/dotnet/sdk:8.0	17	swe_patched (17)	1
19a18711b6cd	elixir:1.16	17	swe_patched (17)	1
5cf4a9f1db56	nickblah/lua:5.4-luarocks-debian	14	swe_patched (14)	1
7d01a21eadb6	python:3.10-slim	10	python_unit_tests (10)	1
aac29b020926	ocaml/opam:debian-12-ocaml-5.1	4	swe_patched (4)	1

By mode

mode	tasks
judge	379355
math	219315
ifeval	46391
json-schema	42798
pytest	40975
stdio	37008
script	27956
mcq	23860
xml-elements	14013
reasoning-gym	13712
exact	13663
csv-columns	2802

By tag

tag	tasks
nemotron	675661
judge	379355
qa	296289
openqa	272429
reference	272429
math	216265
science	150468
code	126064
knowledge	121961
rubric	105874
no-reference	105874
instruction-following	56610
structured-outputs	50446
ifeval	46391
safety	44884
json-schema	42798
stackexchange	40997
competitive-programming	37008
stdio	37008
reasoning	33706
python	28255
swe	26279
swe-repo	26279
mcq	23860
unit-test	22783
kata	22783
arc-agi	19994
json	15652
yaml	14509
xml-elements	14013
xml	14013
reasoning-gym	13712
patched	13559
trusted-test-paths	12720
shell	11497
grid-transform	10000
unix	10000
stackoverflow	10000
superuser	10000
codereview	10000
general-assistant	9999
grid-match	9994
code-assistant	9994
codeforces	9697
structured-output	9167
tool-use	8372
calendar	8372
scheduling	8372
state-tracking	8372
code-contests	8222
script-fallback	8087
puzzle	6719
laion	6719
taco	5116
go	4086
toml	3470
alphabetical-sorting	3110
csv-columns	2802
csv	2802
rust	1837
special-judge	1744
bash	1497
nl2bash	1497
terminal	1497
dcagent2	1497
multi-turn	1052
checklist	1052
tezos	997
binary-alternation	941
basic-arithmetic	920
advanced-geometry	889
count-primes	884
acre	703
java	660
julia	599
kotlin	446
swift	266
zebra-puzzles	178
aiw	177
tsumego	175
leg-counting	164
number-format	161
string-manipulation	161
knight-swap	161
course-schedule	160
ab	160
syllogism	158
spiral-matrix	156
word-sequence-reversal	156
number-filtering	155
figlet-font	155
codeio	153
letter-jumble	152
decimal-chain-sum	152
modulo-grid	152
calendar-arithmetic	152
polynomial-multiplication	152
manipulate-matrix	151
prime-factorization	151
ransom-note	151
count-bits	150
simple-geometry	150
rotten-oranges	149
gcd	149
family-relationships	148
mini-sudoku	148
word-sorting	148
survo	148
letter-counting	148
game-of-life-halting	147
simple-integration	147
mahjong-puzzle	147
needle-haystack	147
simple-equations	147
string-insertion	147
time-intervals	146
number-sorting	146
bf	146
dice	145
gsm-symbolic	145
pool-matrix	145
cryptarithm	145
game-of-life	144
bitwise-arithmetic	143
palindrome-generation	142
spell-backward	142
lcm	141
circuit-logic	141
fraction-simplification	141
emoji-mystery	140
word-ladder	140
kakurasu	140
sokoban	139
group-anagrams	139
palindrome-partitioning	138
maze	137
isomorphic-strings	137
arc-1d	137
rectangle-count	136
sentence-reordering	136
shortest-path	136
caesar-cipher	135
number-sequence	135
rotate-matrix	134
polynomial-equations	132
color-cube-rotation	132
base-conversion	132
list-functions	131
knights-knaves	131
power-function	130
largest-island	130
n-queens	130
sudoku	129
complex-arithmetic	129
binary-matrix	128
countdown	127
puzzle24	127
products	127
decimal-arithmetic	126
futoshiki	125
chain-sum	124
self-reference	124
jugs	123
intermediate-integration	122
quantum-lock	121
coin-flip	104
string-synthesis	103
string-splitting	102
tower-of-hanoi	61
dart	46
c	40
scala	27
php	18
csharp	17
elixir	17
lua	14
cpp	10
ocaml	4

Verifier rejections by check

check	tasks
empty	2433
gold_leak	990

CalibForge native Harbor addition

This release also includes 5,431 tasks from AweAI-Team/CalibForge at fb1e75441a94b8bb0ced08acd6b59e711704d70a. Select source == 'AweAI-Team__CalibForge' from the train split. These tasks retain their native Harbor graders and have no verifyit contract. The preceding conversion report describes the original Task Trove population. The combined current counts and source provenance are in manifest.json.

CalibForge is CC BY 4.0 according to its upstream dataset card. Its original task files and attribution remain in the task archives. Existing source licenses are unchanged.

Snapshot-safe: No. Native environments reference mutable Docker image tags; native test scripts download bootstrap tools and dependencies at execution time. Accepted without making the source snapshot-safe.

SWE-Gym/SWE-Gym native addition

Includes 2,438 canonical tasks from SWE-Gym/SWE-Gym at bb94ed9e39bbeb96a7fcbfb533b80f25a7fd59cb. Select source == 'SWE-Gym__SWE-Gym' in the train split. License: mit. Canonical SWE-Gym tasks with native test selection and grading; oracle patches remain separate.

Snapshot-safe: No. Native environment builds download unpinned packages and use mutable Ubuntu and uv image tags. The native evaluation script can reinstall project dependencies. Dataset and harness code are pinned.