ALE Tasks β Operation Research: run records
The episodes that ale run produces for the Tasks of AgentsLastExam/ale-tasks-or, one folder per Task version. It pairs with that repository's asset dataset ale-or/ale-tasks-or: the asset dataset holds what a Task is made of, this one what runs of it produced.
No runs are recorded yet; run records are added only from real ale run episodes, never made up.
Layout
tasks/<path>/ the Task's path in ale-tasks-or, e.g. tasks/my_task/
βββ <task version>/ sha256-<64 hex>, the Task version (code tree + material tree)
βββ version.json what that version is made of (below)
βββ <run id>/ the run directory name ale run created under runs/
βββ <episode id>/ one episode directory, uploaded as ale run wrote it
βββ result.json terminal status, named rewards, failure, phase timings
βββ lock.json the RunLock: complete provenance of the attempt
βββ trajectory.json the agent's trajectory (Harbor ATIF), when the agent produced one
βββ verification.json the Task verifier's complete result
βββ llm-judge.jsonl, agent-judge.jsonl Judge records, when a Judge ran
βββ trace.transport.jsonl model calls through the Gateway, token usage
βββ trace.execution.jsonl framework phases and commands
βββ blobs/<kind>/sha256-<digest>.<ext> images, documents and other payloads the records reference
βββ attempts/<n>/ earlier attempts an automatic retry replaced, kept as written
What each record owns is the engine's
trace specification; this
dataset changes nothing in them. Consumers read the immediate episode folders (*/result.json) of
a run, never the archived attempts, which are not extra samples.
version.json records the pairing the repository itself does not keep:
{
"task": "tasks/<path>",
"repository": "AgentsLastExam/ale-tasks-or",
"commit": "<repository commit the runs used>",
"code_tree": "<git tree id of tasks/<path> at that commit>",
"dataset": "ale-or/ale-tasks-or",
"dataset_revision": "<asset dataset commit the runs used>",
"material_tree": "<git tree id of tasks/<path> on the asset dataset>",
"version": "sha256:<β¦>"
}
A run belongs under a version only when its RunLock's Task source and asset commit match that version; runs of different versions are never mixed. Results are compared within a version.
Adding a run
Upload one run's episode folders under the Task version they ran, after checking that every
episode's lock.json records clean (not dirty) assets at a synchronized commit:
hf upload ale-or/ale-tasks-or-runs \
runs/<run id> tasks/<path>/<task version>/<run id> --repo-type dataset
Every file under tasks/ is stored with Git LFS (tasks/** filter=lfs in .gitattributes).
Access
Public and gated with manual approval: anyone sees the file list; downloading needs an administrator's approval. Run records contain what the agent saw and produced, the verifier's detailed results and Judge reasoning, so they are never opened without approval. ALE redacts the credentials it knows from every record; review a run before uploading it all the same.
- Downloads last month
- 42