sft-benchmark-trajectories / docs /dataset-cards.md
Heisenger's picture
Publish incremental trajectory release v0.4.0
cf0d2b5 verified
|
Raw History Blame Contribute Delete
2.81 kB
# Adapter dataset cards
The source registry and [coverage inventory](trajectory_sources.md) are
authoritative for revisions and counts. These short cards explain semantic
mapping choices.
## Software-engineering conversations
Adapters: `swegym_openhands`, `swegym_sampled`,
`nvidia_open_swe_traces`, `swebench_verified_trajectories`,
`swesmith_trajectories`.
System/user/assistant/tool messages and native instance IDs are preserved.
String-encoded NVIDIA messages/tools/metadata are JSON-decoded losslessly.
mini-SWE-agent call IDs are namespaced only when the source reuses an ID; the
original value remains in metadata. A final unresolved submit call may be a
valid terminal action when the source publishes its execution verdict, but no
tool result is fabricated. Gold patches remain raw provenance and are excluded
from SFT `model_input`.
## Step-shaped SWE and terminal traces
Adapters: `swebench_multilingual_trajectories`,
`terminalbench_trajectories`, `terminaltraj`.
Every source action becomes a supervised assistant event and every published
response/state/terminal output becomes its causal environment event.
Terminal-Bench trials with `steps=null` are rejected before canonical import;
short or inadequately contextualized traces remain tier-5 diagnostics. Source
reward, cost, duration, model and scaffold remain independent metadata.
## Tool and user-simulation environments
Adapters: `agentsuite_tau2`, `official_tau3`, `toolbench`.
Calls and results remain structured and linked. AgentSuite τ² correctness uses
the published evaluator fields. Official τ³ trials preserve reward evidence,
seed and trial and are protected evaluation data. ToolBench uses only the last
cumulative `train_messages` history so prefixes are not duplicated; its
`win`/`valid_data` flags are not treated as independent semantic proof.
`agentsuite_bfcl` is a diagnostic adapter: it proves the 115,466-row release
contains final function-call answers and evaluator scores but no returned tool
observations. These records fail the trajectory gate and are excluded from the
final library.
## Web interaction
Adapters: `mind2web`, `browsecomp_rollouts`.
Mind2Web pairs each human operation with its full pre-action DOM; it has no
binary task verdict. BrowseComp preserves research-agent search calls, full
returned content and the source answer-judge verdict. Reference answers and
judge explanations stay outside model input. BrowseComp rollouts are protected
evaluation data.
## ALFWorld
The adapter preserves task/scene data, human language annotations, PDDL
high-level actions, low actions and image references. The audited release does
not publish action-paired observations in the planner record, so mapped samples
are explicitly incomplete and are not included in the qualifying library.