Datasets:
TermiGen: stages 1, 3, 5 and teacher results; 3,369 retained tasks
No gold/reference-solution checks were run: TermiGen provides no gold solutions. Broken tasks will likely remain in this dataset and will be filtered out in the next version. Teacher passes are evidence of verifier acceptance, not proof that a task or its tests are correct.
TermiGen_phase-1-3-5-6-7
Keeps 3,369 tasks and archives 197 from ucsb-mlsec/terminal-bench-env, covering all 3,566 source tasks. This PR includes the task payloads, archived tasks with their reasons, and the supporting validation and teacher reports.
The earlier stage-3 publication (PR #8) retained 3,506 tasks and archived 60. This PR adds static and no-op results, excluding a further 137 tasks: 133 with absolute-path findings, 2 with CPU-count findings, and 2 whose verifier passed with no answer. No previously archived task is restored.
Filtering starts with all 3,566 source tasks and applies stages 1, 3 and 5 in order. Each excluded task is counted once, at its first applicable exclusion. The counts use the full validation run and recorded environment reruns.
| Check | Tasks entering this step | Newly archived | Tasks retained after this step |
|---|---|---|---|
| 1: Static checks, with the recorded advisory exceptions | 3,566 | 141 | 3,425 |
| 3: Environment build, startup and inspected files | 3,425 | 54 | 3,371 |
| 5: No-op verification | 3,371 | 2 | 3,369 |
Only 3,336 retained tasks passed all three measured validation phases. The remaining 33 are retained as unresolved under the publisher's existing policy: 28 verifiers executed no tests and 5 verifiers timed out. These are not passing no-op results. Stage 4 was not run because there are no reference solutions.
Static results use the recorded training-profile exclusions; pip-pinning and test-file-reference findings remain advisory. The detailed policy and per-task findings are preserved in the run record.
Archived tasks
The first applicable exclusion accounts for each task once: 141 at stage 1, 54 at stage 3, and 2 at stage 5. Stage 3 still has the same 60 archived outcomes as PR #8; six of those tasks are counted at stage 1 in this combined report.
| Archive label | Tasks | Explanation |
|---|---|---|
| build-command-unavailable | 1 | The build could not find a required shell command. |
| container-start-failed | 7 | Apptainer could not start the environment; the saved logs include fakeroot startup diagnostics. |
| copy-destination-collision | 2 | Build-context copying collided with a file already present at the destination. |
| dependency-build-failed | 12 | Environment setup or dependency installation failed. Per-task errors are preserved in the linked stage report. |
| environment-check-incomplete | 9 | Environment startup or inspection did not finish successfully; the report contains no error text. These tasks remain archived under the recorded phase 3 policy. |
| image-preparation-failed | 3 | CUDA, Lean and Ray image builds failed with dpkg package-configuration errors in all three reruns. Later attempts encountered missing images and did not rebuild them. |
| missing-copy-source | 6 | The Dockerfile references a build-context file that is absent from the task payload. |
| no-op-reward-one | 2 | no-answer reward [1.0] instead of 0 |
| runtime-file-tree-mismatch | 12 | Expected task files were missing or differed from their recorded contents inside the running environment. |
| static:compose-host-binds | 3 | static checks failed: check-compose-host-binds.sh |
| static:nproc | 3 | static checks failed: check-nproc.sh |
| static:task-absolute-path | 135 | static checks failed: check-task-absolute-path.sh |
| unsupported-compose-sidecars | 2 | The Apptainer backend cannot run the required Docker Compose sidecars; all three reruns remained blocked. |
The eight previously unresolved environment tasks remain archived after their recorded reruns: five require unsupported Compose sidecars and three failed image preparation. Later attempts on those three encountered missing images rather than rebuilding them. Nine other environment checks have incomplete startup or inspection evidence and remain archived under the reviewed stage-3 policy. These outcomes do not establish whether a valid solution exists.
Repairs
Our patcher modified 2,219 of the original 3,566 tasks. A task can have multiple changes, so the counts below overlap.
| Change | Tasks affected | Explanation |
|---|---|---|
| deferred-helper | 4 | Reconstruct supplied build helpers when deferred installation runs before COPY payloads are available. |
| env-expansion | 51 | Resolve supported Docker ENV substitutions so inherited command paths remain usable. |
| initial-file | 3 | Materialize initial files from supplied payloads so they remain visible after container mounts. |
| tmp-fixture | 158 | Relocate supplied fixtures hidden by temporary-directory mounts and update their path references. |
| wildcard-copy | 1 | Expand build-context wildcard copies into explicit file copies. |
| workspace-copy | 2,058 | Repair COPY destinations and preserve renamed files through byte-identical aliases. |
The patcher retained all source tasks, including 11 for which no safe automatic patch was available. Their patch predictions and validation outcomes remain recorded separately.
Teacher results: pass@k
Both models used Terminus 2 on the same 3,336 tasks, with four attempts per task and model (13,344 attempts each; 26,688 total). Stages 6 and 7 are informational and do not exclude tasks. The 33 tasks with unresolved no-op results were not included in teacher generation.
| Model | Pass@1 | Pass@2 | Pass@3 | Pass@4 | Tasks solved at least once |
|---|---|---|---|---|---|
| Qwen3-Coder-30B-A3B-Instruct | 48.46% | 57.42% | 61.38% | 63.64% | 2,123 / 3,336 |
| Qwen3-30B-A3B-Instruct-2507 | 37.54% | 47.41% | 52.33% | 55.46% | 1,850 / 3,336 |
Pass@k estimates success within k attempts from the four recorded attempts per task. Missing rewards count as unsolved in these rates; the underlying missing rewards remain missing. At k=4, the rate is the fraction of tasks solved at least once. The models solved 2,277 distinct tasks between them, including 154 solved only by the Instruct model; that combined coverage is not a single-model pass@k result.
| Model | Agent timeouts | Missing rewards | Missing traces |
|---|---|---|---|
| Qwen3-Coder-30B-A3B-Instruct | 414 | 142 | 32 |
| Qwen3-30B-A3B-Instruct-2507 | 1,816 | 179 | 32 |
All planned result records are present, but some rewards and traces are missing. Timeout and execution-error findings are included in the results. The Instruct run completed all 17 stage-6/stage-7 batches within their 4-hour limits. Its Slurm failure statuses reflect recorded trial findings, not interrupted dataset coverage.
Runtime and images
Environment validation used Apptainer on Helma. The original validation evidence records Python 3.12.7 and Harbor 0.8.1; the later eight-task environment reruns used Python 3.12.14, Apptainer 1.5.4, 48 CPUs, 64 GiB per job and concurrency 2. These are historical runs, not a fresh full-dataset reproduction.
The Instruct teacher run used one H200 and 32 CPUs per job, concurrency 16, a 65,536-token context and a 16,384-token reply limit. Jobs took 2h47m to 3h29m and consumed 53.57 GPU-hours in total. The Coder run used concurrency 10 and consumed 70.24 GPU-hours; model and concurrency both changed, so this is not an isolated concurrency-speedup comparison.
Existing cached images were reused for teacher generation. No cached image bundles are included in this PR because the legacy cache lacks the release provenance required by the current publisher. Task payloads and Dockerfiles are included.
Files and evidence
The files below are included in this PR; the patch-script link points to the recorded GitHub commit. The teacher reports contain summaries and per-attempt metrics; the raw conversations and trial logs remain in the original run archives referenced by the provenance files. The image manifest records missing release bundles; it does not contain runnable images.
- Retained tasks and archived tasks with reasons.
- Run record and per-task findings.
- Validation reports: stage 1, reviewed stage 3, and stage 5.
- Teacher comparison and pass@1 through pass@4.
- Instruct results: stage 6, stage 7, and run summary.
- Coder outcome and trace summary.
- Input evidence and teacher-batch provenance.
- Patch comparison with upstream, patcher reproduction check, and patch script at the recorded commit.
- Image release manifest.