sft-benchmark-trajectories / docs /real-data-smoke.md
Heisenger's picture
Publish incremental trajectory release v0.4.0
cf0d2b5 verified
|
Raw History Blame Contribute Delete
3.11 kB

Real-data smoke tests

These commands are retained for small adapter smoke checks. The full release workflow is now scripts/import_hf_release.py; exact revisions, complete counts and commands are documented in trajectory_sources.md.

Store downloaded source payloads in the ignored data/cache directory or an external cache. Only canonical records that pass the strict trajectory gate are persisted under data/real.

The helper below downloads metadata, split information and complete rows from the dataset-server rows endpoint; it rejects truncated cells, records the immutable dataset revision and file hash, and can then exercise the matching adapter without downloading an entire corpus.

python scripts/audit_hf_source.py \
  swegym_openhands SWE-Gym/OpenHands-SFT-Trajectories \
  --cache /workspace/shared/trajectory-data --adapt

python scripts/audit_hf_source.py \
  nvidia_open_swe_traces nvidia/Open-SWE-Traces \
  --cache /workspace/shared/trajectory-data --adapt

python scripts/audit_hf_source.py \
  terminaltraj m-a-p/TerminalTraj \
  --cache /workspace/shared/trajectory-data --adapt

python scripts/audit_hf_source.py \
  agentsuite_tau2 AgentSuite/tau2-bench-trajectories \
  --cache /workspace/shared/trajectory-data --endpoint first-rows --limit 5 --adapt

python scripts/audit_hf_source.py \
  mind2web osunlp/Mind2Web \
  --cache /workspace/shared/trajectory-data --limit 2 --adapt

If a dataset has multiple configurations or splits, pass --config and --split after inspecting the helper's first response. An adapter schema error is an audit finding: update the source audit and adapter from the actual row; do not coerce or invent fields.

For the two GitHub-accessible smoke tests reproduced during onboarding:

trajdata adapt official_tau3 /path/to/tau2-bench/data/tau2/results/final/RESULT.json \
  --output /tmp/tau3-canonical.jsonl \
  --revision dd8757d4c8791dad7b3cdaf02a105a3d5f9d86bd \
  --split audited-file
trajdata validate /tmp/tau3-canonical.jsonl

trajdata adapt toolbench /path/to/ToolBench/data_example/answer/G1_answer/10_ChatGPT_DFS_woFilter_w2.json \
  --output /tmp/toolbench-canonical.jsonl \
  --revision d56fdd89faf8c91fa135090b212bb9057ee5cfc2 \
  --split checked-in-example
trajdata validate /tmp/toolbench-canonical.jsonl
trajdata html /tmp/toolbench-canonical.jsonl --output /tmp/toolbench-gallery.html

curl --fail --location \
  --output /tmp/alfworld-json_2.1.1.zip \
  https://github.com/alfworld/alfworld/releases/download/0.2.2/json_2.1.1_json.zip
sha256sum /tmp/alfworld-json_2.1.1.zip
unzip -t /tmp/alfworld-json_2.1.1.zip
# Extract a selected traj_data.json outside Git, then:
trajdata adapt alfworld /tmp/alfworld-sample/traj_data.json \
  --output /tmp/alfworld-canonical.jsonl \
  --revision release-0.2.2-sha256-25171f16e20ad7b048c47275c45b0babf3aa1cbab29cec97387922350a9844bc \
  --split valid_seen
trajdata validate /tmp/alfworld-canonical.jsonl

Never use the official τ or Terminal-Bench results in --training exports. The registry gate enforces that restriction.