Known limitations
- NVIDIA Open-SWE-Traces is very large. The complete v1.2/OpenHands population is imported, but the other 626,189 rows across v1.0, v1.1 and the v1.2 SWE-agent/mini-SWE-agent splits are not persisted locally. Their source Parquet volume is 55,056,684,304 bytes.
- No benchmark container was independently replayed. Outcomes are clearly labelled as source-evaluator/source-reported evidence, never as local reproduction. This leaves quality tiers 1 and 2 empty.
- No baseline language model, model weights or inference accelerator is configured. The pinned Nemotron tokenizer/chat-template demonstration builds correct labels and batches, but teacher-forced model NLL remains not run.
- The public dataset is federated. It hosts full canonical payloads for ten sources and complete metadata/pinned pointers for three sources whose transformed-redistribution rights remain uncertain. Authorized users can supply locally ingested overlays; access is not automated or implied.
- AgentSuite BFCL contains evaluated final function-call answers but no returned observations. ALFWorld planner files likewise omit action-paired environment observations. They are coverage findings, not trajectory-library records.
- DeepSWE/Datacurve did not yield an exact public, pinned full-trajectory artifact during discovery. This means none was found, not proof that none exists.
- The checked-in ToolBench repository exposes only 15 examples; 11 satisfy the strict trajectory gate. Full upstream generation data was not independently enumerated.
- Mind2Web's hosted Parquet conversion exposes 1,009 records. The conversion is marked partial upstream, so this is complete for the exposed conversion but not claimed as the entirety of every historical Mind2Web artifact.
- Cost, latency and token fields are sparse and source-specific. Raw values are retained, but incompatible units are not pooled into a misleading global comparison.
- Tool definitions are absent from several releases. Calls and returned observations remain usable, while the missing definitions are declared rather than reconstructed.
- Multimodal artifacts are referenced, not embedded or fetched. Raw records can be large because lossless source preservation is intentional.
- Task overlap prevention depends on stable/native IDs and the explicit family normalization implemented by adapters. It cannot discover semantic overlap between unrelated identifiers automatically.
Exact per-source gaps and denominators are in
trajectory_sources.md and
data/coverage/trajectory_sources.json.