SETA v55: 3,083 validated tasks and cached images
#4
by FWeindel - opened
camel-ai__SETA-Env
Keeps 3,083 tasks and archives 70 from open-thoughts/TaskTrove at commit 9262d5628e13ec20ac75b7d897f94b93f6be0594. Patch version: seta-stage1-v55.
| Archive label | Tasks | Explanation |
|---|---|---|
| dependency-setup-too-slow | 6 | Dependency setup exceeded the reviewed 20-minute cutoff with 1 CPU and 2 GiB RAM. Six tasks remain archived; these are setup decisions, not incorrect-reward claims. |
| patch-drop:backend_build_ownership | 19 | The current Apptainer builder cannot preserve the root:ssh_keys ownership required by these Rocky Linux tasks. |
| read-only-nvidia-tool | 2 | Setup tries to replace /usr/bin/nvidia-smi, which is mounted read-only on the validation nodes. |
| static:check-test-file-references.sh | 43 | The checker found files expected by tests that were not established by the instructions or supplied task files. Archived pending review. |
Our patcher modified 3,052 of the original 3,153 tasks.
| Change label | Tasks affected | Explanation |
|---|---|---|
| environment-adapted | 215 | Update environment build files for the repaired tasks. |
| instruction-clarified | 25 | Clarify required paths, arguments and expected outputs. |
| reference-solution-updated | 14 | Repair reference execution and align it with task requirements. |
| setup-updated | 72 | Repair task setup and pin its dependencies. |
| task-configuration-updated | 42 | Adjust setup timeouts and memory for tasks reviewed during validation. |
| verifier-updated | 3048 | Pin dependencies, improve installer error handling, repair file and output checks, and isolate SSH verification. |
Publication also appended or updated the required instruction suffix. This changed the 82 tasks otherwise untouched by the patcher, bringing the final number of tasks differing from upstream to 3,134.
Runtime and validation time
| Setting | Recorded value |
|---|---|
| Cluster | Helma |
| Backend | Apptainer |
| Apptainer (original full run) | 1.5.4-1.el9 |
| Validation Python | 3.12.14 |
| Harbor | 0.8.1 |
| Architecture | x86_64 |
| Network | Host networking |
| CPUs (original full run) | 32 |
| Terminal-Bench commit | 1dcda8716784493721921c23e4bc7f7d988b4494 |
| Harbor commit | 7faf878c14b6d72737579ec936ebf5f74ba3194d |
| Unique cached images / tasks kept | 11 / 3,083 |
These timings cover the original 3,130-task run (Slurm job 934523). Later repair and retry jobs are not included.
| Stage | Node | Tasks | Tasks at once | Duration | Median per task | 90th percentile |
|---|---|---|---|---|---|---|
| 1 | h24-01 | 3130 | 32 | 0:11:52 | 0.7 s | 2.2 s |
| 3 | h24-01 | 3130 | 16 | 1:57:15 | 28.4 s | 51.7 s |
| 4 | h24-01 | 3130 | 16 | 3:20:44 | 35.6 s | 132.0 s |
| 5 | h24-01 | 3130 | 16 | 4:05:17 | 38.4 s | 167.2 s |
Total recorded stage time: 9:35:07. Excludes queue time and work outside these stages.
Static checks excluded by the contract: 14. Recorded but not required: none. Reasons are in the run file.
Image manifest: images/manifests/2026-10-07-seta-v55-reviewed.json. Cached images use Apptainer; Dockerfiles remain in the tasks.
Automatically generated data card:
Validation details (camel-ai__SETA-Env)
- The four nproc tasks are restored: their reviewed commands reported the allocated 1 or 2 CPUs on Helma. These exceptions do not establish Docker quota-only portability.
- The v55 SSH verifiers use per-trial pipes and process-group cleanup, removing shared port 2222 and broad pkill interference. Under fakeroot the server uses a nested user namespace with a non-root identity. Both reference solutions score 1 and both no-ops score 0 in concurrent ten-task pilot 948484; all ten pass stages 1, 3, 4 and 5. The verifier requires a real handshake and confirms the original binary displays the banner.
- Six slow setups remain archived. Completed setup measurements were 20m06s for the desktop task, 25m39s for the lung-cancer R task, and 38m57s for the regression b1__d1 task, whose glmnet compilation was killed at the memory limit. Three other R setups exceeded 40 minutes and were cancelled before completion. Measurements used 1 CPU and 2 GiB per task.
- Published Apptainer images cover Dockerfile builds. SETA still runs its per-task setup script when starting a fresh task; cached images do not preinstall every task-specific dependency.
- Stages 2 and 6 onward were not acceptance gates. No teacher pass@k or LLM rubric results are claimed. Absolute-path syntax was excluded in the original run and checked only in a small subset of later trials. AI detection was skipped without a GPTZero key. Host networking does not establish offline isolation.