A provenance-preserving release of 174,259 upstream software-engineering agent trajectories, with an audited default selection of 11,699 deduplicated tasks, their complete structured conversations, and 100 native training parts. Failures, excluded candidates and duplicate lineage remain available separately. The release combines Microsoft Orchard SWE and Nebius SWE-rebench OpenHands trajectories.

source_success records the upstream publisher's outcome. It does not mean that LANA-research freshly reran that trajectory's patch. The source-bank field execution_reverified is false. A terminating tool call, an agent's success claim or a successful shell exit is insufficient to set source success.

The repository is intended to be public with manual download approval. Gating governs initial access; it does not replace the licenses or add restrictions on rights already granted upstream. Publication does not alter an existing production training manifest.

Configurations and counts

Configuration	Rows	Meaning
`retained` (default)	11,699	Source-resolved, validated, deduplicated selection with known base commit and complete native full fit
`duplicate`	59,687	Validated source-positive attempts linked to the final canonical retained task
`negative`	67,449	Original source outcome is unsuccessful; labels and failure evidence are retained
`quarantine`	35,424	Source-positive attempts excluded by structural, argument, heuristic or native admission checks
`catalog`	174,259	Union of the four configurations above
`native`	11,699	The retained selection as native token vectors and assistant target spans, in 100 parts
`raw_orchard`	107,185	Orchard SWE schema and publisher labels; individually recorded publication substitutions preserve row count and types
`raw_nebius`	67,074	Nebius schema and publisher labels; individually recorded publication substitutions preserve row count and types
`qwen_diagnostics`	12	New observed model task attempts across 10 distinct executable tasks; no positive training admissions
`qwen_preparations`	7	Environment/protocol preparations with no task-model inference
`qwen_calibrations`	3	Protocol calibration responses; these are not task attempts

canonical_instance_id is a convenience alias: validated ingestion events use the verified upstream alias; rows without such an event fall back to the original instance_id. It is not an independently canonical task key for all 174,259 attempts. Retained task counts use the final canonical_id deduplication root together with repository/base-commit identity. Negative or quarantined attempts are not claimed to be independent tasks.

catalog already includes all four status subsets. native and its length cohorts refer to the same retained tasks; raw-source configurations refer to the same original trajectories represented by catalog. Do not concatenate overlapping configurations and count them as new data. raw_orchard and raw_nebius have different original schemas and are intentionally separate. This is a training/analysis release, with no newly constructed validation or test split.

The retained source attribution is Orchard 8,348 and Nebius 3,351 tasks. Teacher counts are MiniMax-M2.5 8,032, Qwen3.5-397B-A17B 316, and Qwen3-Coder-480B-A35B-Instruct 3,351. These identify upstream generation teachers, not the tokenizer used for the native export.

Original and structured records

All catalog configurations use the same publication schema. Important columns include:

record_id, status, canonical_id, reasons: stable trajectory identity, final disposition and exclusion evidence. A trajectory is not an independent task; multiple successful attempts may share a canonical task.
source_dataset, source_revision, source_file, source_row, source_file_sha256: exact upstream provenance. source_file_sha256 identifies the original upstream file; the public release manifest independently identifies published file bytes, including any explicitly recorded public source projections.
source_success, execution_reverified, teacher, repo, instance_id, canonical_instance_id, base_commit: outcome qualification and task identity.
messages: a typed convenience projection for viewers. It can insert null placeholders for absent optional fields and is not the authority for unknown message attributes.
original_source_json: the full source row as canonical JSON, with the explicitly documented publication substitution below. Source JSON string/list/object/null types and tool-call argument strings are retained. For unchanged rows, its canonical logical digest is original_source_sha256; affected rows separately record original and public projection digests in the redaction audit. These are distinct from Parquet file byte digests.
tools_json: a decoded-tools convenience view serialized as JSON. Consult original_source_json for the original field type: Orchard's tools and metadata are JSON strings; Nebius's tools is a list. Original tool_calls[*].function.arguments strings remain strings. Missing, empty and null values are not interchangeable.
record_json: publication metadata, the final task/outcome/validation/reason/lineage records, and native bindings. Ingestion events may precede transitive deduplication; final status and canonical identity are authoritative.
native_file, native_row, cohort, native_tokens, assistant_tokens, view_tokens, views, fullfit: exact retained-native bindings and measurements. Unmeasured excluded rows use null counts; null does not mean zero tokens.

Candidate model_patch is agent-generated output. Source quality fields such as gen_tests_correct and pred_passes_gen_tests are preserved with their original numeric/null types; they are not substituted for the original resolved outcome. They do not contain the hidden reference patch.

import json
from datasets import load_dataset

row = next(iter(load_dataset(
    "LANA-research/LANA-Code-Agent-Trajectories", "retained",
    split="train", streaming=True, token=True,
)))
raw = json.loads(row["original_source_json"])
messages = raw.get("messages", raw.get("trajectory"))
original_tools_value = raw["tools"]
metadata = json.loads(row["record_json"])

Admission and deduplication

The frozen upstream pool contains 106,810 source-resolved and 67,449 source-unsuccessful attempts. The positive bank uses only the former. Task joins use pinned upstream task identity, repository and base commit. Strict indexed keys cover original source identity, task identity, transport-normalized question text and qualified nonempty observed patches where available. Different known base commits remain distinct. Similarity and semantic equivalence beyond these keys are not claimed. Patch identity coverage is partial; an absent or unqualified patch is never hashed as an empty common patch.

Tool definitions, unique call/result bindings, temporal order, closure and the supported required/type/enum/object/array/composition argument constraints are checked. An exact declared finish call followed immediately by its bound message echo is accepted as a closed terminal acknowledgement. The original acknowledgement remains in the conversation; no assistant message is fabricated. Unknown/nonterminal endings remain excluded. Source-positive labels never override malformed tools, failure flags or native-fit exclusions.

Leak/history/network and repetition screening is heuristic and conservatively excludes flagged candidates. It is not proof that every possible contamination or secret is absent. The license and publication screening is separate from training admission. All original dispositions and final transitive lineage remain available.

Native tokens and loss masking

The published native export contains 522,651,820 complete conversation token IDs, 176,550,036 assistant target tokens, and 563,415 assistant training views. Repeated-prefix view lengths total 16,726,943,021 tokens; that total counts reused history and is not a count of new conversations or unique text.

The unchanged private parent contained 522,655,356 conversation IDs, 176,553,041 assistant targets, 563,415 views and 16,726,976,292 repeated-prefix view tokens. Original and published vector/file digests are separate namespaces; the figures above describe the actual public vectors.

Stored full-fit cohort	Tasks
8,192	8
16,384	493
32,768	3,640
65,536	5,222
131,072	2,306
262,144	30

Each native row has record_id, input_ids: list<uint32>, assistant_spans: list<{message_index,start,end}>, and cohort. Span ends are exclusive. A logical view for one assistant span uses input_ids[:end] and predicts only input_ids[start:end]; earlier context is conditioning with zero loss. User, tool and environment frames are not targets. Future observations, including a terminal tool acknowledgement, are excluded from that assistant view. Do not train every position of the stored full vector with positive loss.

The SLF physical packing rule is P + 2*T + 1024 <= cohort, using the maximum complete view per trajectory. It is not plain-text length or the Qwen API token counter. The audited reader implements the recorded serialization, masking and packing; no trajectory or assistant target is silently truncated.

The fixed tokenizer is the exact official NVIDIA tokenizer at revision a1cba248ba7a6a9905380a8622b81f532fcb4913, SHA-256 c6021eb6847e682f89aa52d5eb6e8c7d902a23acfc8137e25211cf84828f1592. The wire is native_chatml_public_tool_json_v1; its end marker is <|im_end|> / token ID 11. A legacy model configuration's EOS ID 2 is not this wire's end marker. Model weights and model configuration are not part of this release. The tokenizer has its own NVIDIA license and notice.

Fresh Qwen diagnostics

The optional qwen_diagnostics, qwen_preparations and qwen_calibrations configurations are separate diagnostic records, not additional retained positives. The frozen smoke cohort covers 10 unique training tasks: 12 actual model task attempts, 7 preparation/transport attempts with no inference, and 3 protocol calibration responses. There were zero new positive training admissions. Failed, unsubmitted and format-invalid attempts keep their original reasons.

The publication keeps actual model responses and reasoning, model-visible tool outputs, candidate patches and recorded official verdicts. Routing credentials, private service receipts, machine-local operational paths and hidden evaluator patch/test payloads are excluded or explicitly recorded as publication redactions. Per-test names/statuses/counts may be retained in verdicts; these are outcome evidence, not hidden test implementation. Calibration is not benchmark success. A public artifact hash is distinct from an original artifact hash when a targeted redaction was required.

Integrity, provenance and rights

The original bank passed full source-label, lineage, native-wire and assistant-span reconciliation over 4,457 ingest shards and 22 source files. The 100-part compact export was independently reread against every selected parent vector/span digest. The public release manifest records each published file's bytes and SHA-256. Portable public metadata has its own digest and identifies the original private artifact digest; a portable metadata rewrite is not described as a byte-identical copy. Source Parquet and native parts are byte-identical only where their exact-copy pins say so.

Publication screening conservatively selected material at 208 exact source review locations across 76 trajectories for substitution. The public copy replaces 216 bounded physical source spans, including 71 additional alias-only display spans of the same frozen material identities, and their recorded structured/historical mirrors. A raw JSON observation, its decoded child and a mirrored event are different counting units; these figures are not a count of distinct credentials. Source outcome, task identity, teacher and disposition are unchanged. 14 of the 22 upstream files are documented public projections with their original schemas and row counts; 8 remain byte-identical. All 11,699 retained task identities and cohort memberships remain unchanged; 7 retained vectors in 6 of the 100 training parts are re-encoded from the projected structured source, with every assistant prefix/target/mask checked. The remaining 94 native parts keep their original bytes.

The redaction map records exact record/field locations, occurrence counts, material hashes and original/public digest namespaces. Elision controls and counts, unrelated public keys, test names, parameters and code remain intact. The immutable private parent bank is unchanged. These source-bank publication projections have not been re-executed; publisher verdicts refer to the original trajectories. Publicly verified fixtures, synthetic marker strings and documented transient task-service logs are preserved with exact safety bindings. Unresolved material is not described as a verified fixture. Some publicly documented keys and test vectors share components with the projected material. These counts describe the reviewed substitutions, not distinct unresolved credentials; a shared component does not establish the provenance of an entire wrapper.

The original Orchard and Nebius license terms are retained in LICENSE, THIRD_PARTY_LICENSES.md, NOTICE, licenses/ and source attribution records. Orchard declares MIT subject to the licenses of the individual target repositories; Nebius declares CC BY 4.0. Repository code, patches and tests keep applicable repository rights. A trajectory dataset license does not relicense every embedded code excerpt. The release does not claim that every represented repository's license has been individually adjudicated. NVIDIA's tokenizer license is a separate component license. No endorsement by Microsoft, Nebius, NVIDIA, MiniMax or Qwen is implied.

The source tasks and traces are unsuitable as an unbiased held-out benchmark. Review provenance, training overlap and the applicable source/repository licenses for the intended use. See the pinned source cards and CITATION.bib for attribution.

Reproducibility code and ingestion archives

The portable reader, wire implementation and frozen CPU emitter distributed under code/ are LANA-research packaging code licensed under MIT. Their license does not replace the upstream dataset, target repository or NVIDIA tokenizer licenses. No original trainer implementation or production training manifest is included.

An optional ingestion archive preserves historical shard metadata, intermediate disposition evidence and native exports, including rows later demoted by deduplication. Its original status and canonical identifiers are historical; the released catalog and final lineage supply the authoritative final disposition. 24 gzip members and 6 historical native members are documented public projections. Only recorded physical material spans, mirrored representations and their derived vectors change; all other JSONL lines and archive member bytes are preserved. The archive index records original and public digests for transformed members and repacked outer tars. Unmodified historical shard receipts describe the original file-hash namespace; use the publication map and public archive index for current bytes. Outer tar headers use portable zeroed identity/time fields. Archive membership and digests are separate from the compact training manifest. Only final retained native rows are eligible for the default training selection.

Downloads last month: 22

Total file size:

17.5 GB