docs/trajectory_sources.md

Trajectory source inventory

Audit date: 2026-10-14. Counts below distinguish upstream rows, imported canonical records and records passing the strict multi-action/context gate. data/coverage/trajectory_sources.json is the machine-readable counterpart; data/library/manifest.json hashes every local canonical shard.

“Complete import” means the full identified upstream population for that release/split, not every artifact ever published for a benchmark. “Protected” means usable for evaluation/inspection but blocked from training export. Source outcomes were not independently replayed unless explicitly stated.

Release note: the detailed sections below document the original v0.3.0 build. The append-only v0.4.0 release expands the inventory from 155,278 to 235,876 trajectories without rebuilding those records. See the v0.4.0 incremental audit for its pinned sources, accepted/rejected counts, rights decisions and exact append command.

Measured coverage

Imported counts successfully canonicalized source rows before exact deduplication. Canonical collapses only records with the same causal interaction and the same outcome/reward/completeness/verification evidence; all contributing IDs and locators remain in duplicate_provenance.jsonl.gz. S/F/U are canonical success/failure/unknown counts. Strict is the retained canonical count passing the complete multi-action/context gate.

Source	Upstream rows	Imported	Canonical	Unique tasks	Strict	S/F/U	Missing or collapsed
SWE-Gym OpenHands SFT	491	491	491	293	491	491/0/0	none
SWE-Gym Sampled	6,055	6,055	5,502	2,438	5,502	491/5,011/0	553 identical records collapsed
SWE-bench Verified trajectories	5,000	5,000	4,544	500	4,532	3,392/1,152/0	456 identical records collapsed; 12 retained partials
SWE-bench Multilingual trajectories	129	129	129	43	110	8/121/0	19 retained partials
SWE-smith tool	24,100	24,100	24,100	10,637	24,100	9,427/14,673/0	alternate XML/ticks serializations excluded
NVIDIA v1.2/OpenHands	677,815 all-release	51,626	51,626	33,444	51,626	24,743/26,883/0	626,189 other version/scaffold rows not imported
Terminal-Bench trajectories	52,104	34,462	34,263	89	24,086	11,051/23,212/0	17,642 have no steps; 199 identical records collapsed
TerminalTraj	20,000	20,000	20,000	17,787	19,143	0/0/20,000	857 retained short/partial traces
AgentSuite τ²	8,340	8,340	8,340	228	8,337	3,375/4,965/0	3 retained incomplete traces
Official τ³ checked result	200	200	200	50	0	100/100/0	protected; not SFT-complete under strict gate
BrowseComp rollouts	5,064	5,063	5,063	1,266	5,063	1,211/3,852/0	1 malformed row
Mind2Web exposed conversion	1,009	1,009	1,009	1,009	1,009	0/0/1,009	binary task outcome absent
ToolBench checked examples	15	11	11	11	0	7/4/0	4 fail final gate; source flags are not semantic proof
AgentSuite BFCL	115,466	0	0	0	0	0/0/0	final calls only; no observations
ALFWorld planner records	7,080	0	0	0	0	0/0/0	actions lack paired observations

Across the final inputs, 156,486 source rows became 155,278 canonical interactions after 1,208 exact duplicates were collapsed. They cover 67,610 normalized tasks; 143,999 interactions pass the strict gate and 110,208 are also training-eligible after protected-source separation. The coverage manifest additionally records official final-answer/task registries and the DeepSWE search result that do not enter this table as trajectory releases.

SWE-Gym and SWE-bench family

• SWE-Gym OpenHands SFT, revision 4aaa5a4a4b5861f4799d2336908760c190ac3b17, MIT: all 491 train.success.oss OpenHands completion trajectories imported. The success label is the published split, not local replay. Messages encode tool actions/results in content; no arbitrary sample cap was used.
• SWE-Gym OpenHands Sampled, revision baf3a4e4bff514d48ddc08a93a2ade5c126212c7, no card license: all 6,055 attempts imported, including published successful and failed execution outcomes, structured tools and reports.
• tarsur385 SWE-bench Verified trajectories, revision 773748a7c1222e8a642a7059821498e14293562a, MIT: all 5,000 mini-SWE-agent files (10 model families × 500 protected tasks) imported. Calls, results and source-resolved verdicts are full trajectories; all remain protected evaluation data. The official SWE-bench Verified instance table contains issues/tests/gold patches, not trajectories.
• AlienKevin SWE-bench Multilingual trajectories, revision 7221c09c6cc0ed16d51c9cb9886a1ab484465670, MIT: all 129 attempts (three models × 43 Rust tasks) imported; 110 pass the strict contextualized multi-action gate. The remaining 19 are explicit tier-5 partial/short diagnostics. This is a third-party protected rollout release, not an official expert-demonstration set.
• SWE-smith trajectories, revision 08e109b4a59eaeebf80e4675cd125d42e7ac99a4, MIT: all 24,100 tool rows imported and accepted. The xml (26,076 rows) and ticks (25,826 rows) configurations are alternate serializations rather than additional independent rollouts, so they are deliberately not duplicated.

NVIDIA Open-SWE-Traces

nvidia/Open-SWE-Traces was pinned at 6d0a83373ae0d83abe3d363f1f18cb1822e16586 (CC-BY-4.0). The release enumerates 677,815 trajectories:

Version	Rows
v1.0	151,219
v1.1	263,226
v1.2	263,370
All	677,815

All 51,626 v1.2/OpenHands records across 24 revision-qualified source Parquet shards were imported and passed the action/observation gate. Messages, tools and metadata were JSON-decoded; native instance/model/scaffold/resolved fields remain intact. The other 626,189 version/scaffold rows were not downloaded: their source Parquet totals 55,056,684,304 bytes, which exceeds the remaining workspace budget once lossless canonical records and indexes are retained. This is an explicitly quantified partial import, not a sample represented as a complete NVIDIA corpus. The upstream repository advanced during the audit, but all imported URLs and metadata remained pinned to the stated SHA.

Terminal agents

• yoonholee Terminal-Bench trajectories, revision 04e8940f5b6736a7ce8d22224fe2f2af74163ed2, Apache-2.0: all 52,104 trial rows enumerated. Exactly 34,462 publish step sequences and were imported; 17,642 steps=null rows are rejected as absent trajectories. Of the deduplicated traces, 24,086 pass the strict gate and the rest remain tier-5 partial diagnostics. Model, scaffold, reward, duration, token and cost fields are preserved. Observations are source-truncated to 5,000 characters. This protected third-party release is distinct from the official task repo.
• TerminalTraj, revision 5c1823f4a8b9ca0cf02c27d2db52c5b35b53a308, no card license: all 20,000 command/output records imported; 19,143 pass the multi-action/context gate. No trustworthy task-level verdict exists, so outcomes are unknown—not failed.
• The official Terminal-Bench repository at d28711d0da2675d0bb1d56de45ae5df6082438a3 contains 241 protected task manifests, tests and reference scripts. Those are not completion trajectories; 966 audited files carried explicit anti-training canaries.

τ² and τ³

• AgentSuite τ² trajectories, revision 07264f1d253a618fc20966670eddc44246212916, no card license: all 8,340 rollouts (30 models × 278 task slots) imported; 8,337 pass the complete action/observation gate and three remain explicit incomplete diagnostics because a call has no returned observation. System/user/assistant/tool interactions and evaluator correctness/reward are present. Tool definitions are absent and declared missing. They are τ² rollouts, not assumed version-compatible with τ³.
• Official τ³ repository rollouts, revision dd8757d4c8791dad7b3cdaf02a105a3d5f9d86bd, MIT code plus benchmark restrictions: the available checked-in final result file contains 200 model/user-simulator trials over 50 tasks (100 source successes and 100 failures). It is imported for evaluation only. Expected actions and solution metadata are never treated as expert demonstrations.

Function calling and general tool use

• AgentSuite BFCL outputs, revision a86612c6bb68e6566e5e7d206f7b031af889edc0, no card license: all 115,466 rows were audited. They contain user prompts, model final function calls, scores and latency but no returned tool observations. Zero qualify as trajectories; none enter the canonical library. This is “exists but is only final output,” not “no BFCL data exists.”
• ToolBench, revision d56fdd89faf8c91fa135090b212bb9057ee5cfc2, Apache-2.0 repository: all 15 checked-in answer examples were inspected, 13 adapted, and 11 pass the strict gate. The final cumulative history is used once; cumulative prefixes are not duplicated. win/valid_data are generation flags rather than independent semantic success evidence. A complete full-release population was not verified.

Web agents

• rl-rag BrowseComp rollouts, revision ac4854be0eb527cdf90f6c7d09446eeb7a3d1dbd, no card license: 5,063/5,064 full research-agent conversations imported; one row lacks a conversation. The release normally has four GPT-OSS-120B attempts per question, full search results and a source GPT-4o answer-judge verdict. Reference answers stay out of model input. This is protected evaluation data; official BrowseComp question/answer rows alone are not trajectories.
• Mind2Web, revision 17ece8eb89862368edc0cc806acee6fca5163474, CC-BY-4.0: all 1,009 rows exposed by the hosted partial Parquet conversion imported. These are human action sequences with a full pre-action DOM for each target and no binary task-success field. Completeness is claimed only for that exposed conversion, not every historical artifact.

ALFWorld and DeepSWE

• ALFWorld release 0.2.2 contains 7,080 planner traj_data.json records (archive SHA-256 25171f16e20ad7b048c47275c45b0babf3aa1cbab29cec97387922350a9844bc). They preserve human task text, planner/low actions and image references but not action-paired environment observations in the audited artifact. They are therefore excluded rather than promoted to full trajectories.
• No exact public, pinned DeepSWE/Datacurve full-trajectory release was found. This is the “none found after searching” state—not evidence that no private or future trajectory artifact exists. Protected SWE evaluation/reference data was not substituted for demonstrations.

Outcome and permission interpretation

• Expert provenance is used only for human/reference/planner demonstrations; successful model rollouts remain recorded agent completions.
• Source-evaluator and source-reported verdicts are distinct from independent reproduction. This build performed zero native benchmark replays.
• Conditional sources can populate preparation-time training exports because licensing was explicitly out of scope for this data-preparation task. Evaluation-only sources remain protected_evaluation regardless.
• Failed and unknown outcomes have separate indexes and denominators.

Reproducing imports

Most hosted releases use the same resumable command (repeat --split for multi-split datasets):

.venv/bin/python scripts/import_hf_release.py SOURCE_ID DATASET_ID \
  --revision IMMUTABLE_SHA --output data/full/canonical/OUTPUT \
  --cache data/cache/full/OUTPUT --config default --split SPLIT

Concrete source/split pairs are recorded in data/registry/sources.json and every generated release-manifest.json. NVIDIA uses pinned repository files so the moving dataset-server conversion cannot change the bytes:

.venv/bin/python scripts/import_hf_release.py \
  nvidia_open_swe_traces nvidia/Open-SWE-Traces \
  --revision 6d0a83373ae0d83abe3d363f1f18cb1822e16586 \
  --output data/full/canonical/nvidia_v1_2_openhands_original \
  --cache data/cache/full/nvidia_v1_2_openhands_original \
  --config v1.2 --split openhands \
  --original-glob 'data/openhands/qwen38_27b/*/*.parquet' \
  --original-glob 'data/openhands/qwen38_flash_next/*/*.parquet'

Native JSON trees (tarsur SWE-bench Verified, AgentSuite τ² and official τ³) are checked out at the registered SHA and streamed with:

.venv/bin/python scripts/import_json_tree.py SOURCE_ID PINNED_CHECKOUT \
  --output data/full/canonical/OUTPUT/all.jsonl.gz \
  --revision IMMUTABLE_SHA --url DIRECT_SOURCE_URL --glob '**/*.json*'

The combined 211-record seed used by the final build is reproducible from the official τ³ result and checked-in ToolBench example. First adapt those pinned files separately; apply scripts/filter_canonical.py --strict --source toolbench to the ToolBench result (13 adapt, 11 pass); then combine the 200 τ³ and 11 ToolBench records with:

.venv/bin/python scripts/filter_canonical.py \
  data/full/canonical/official_tau3/all.jsonl.gz \
  data/full/canonical/toolbench/strict.jsonl.gz \
  --source official_tau3 --source toolbench \
  --output data/full/canonical/seed_evaluation/official_tau3_toolbench.jsonl.gz

Finally run the complete-library, coverage, gallery and evaluation commands in the root README. No command applies an arbitrary row limit.