Datasets:
Endless Terminals: 2,145 retained tasks, 347 archived; phases 3–5
Endless Terminals
Keeps 2,145 tasks and archives 347 from obiwan96/endless-terminals, accounting for all 2,492 source tasks. Includes task payloads, archive reasons, native Apptainer images, validation evidence and the measured teacher trajectories.
Filtering applies phases 3, 4 and 5 in order. Phase 1 filtering is disabled; existing dependency, setup and solution repairs are retained. Includes the full gold/no-op validation run and the repaired-task pilot results. Historical validation is reused after an instruction-only change; the published instruction variant has not been rerun.
| Filtering phase | Tasks entering | Newly archived | Tasks retained |
|---|---|---|---|
| 3: Image build/start and reviewed conversion findings | 2,492 | 271 | 2,221 |
| 4: Gold/reference solution | 2,221 | 53 | 2,168 |
| 5: No-op verification | 2,168 | 23 | 2,145 |
Each excluded task is counted once, at its first applicable exclusion. Across all source tasks, 2,434 pass gold and 58 fail; 2,468 pass no-op and 24 fail or repeatedly produce an invalid result. Five gold failures and one no-op failure were already excluded at phase 3.
Conversion findings and archive reasons
All 2,492 task images built successfully. A focused source-based conversion audit inspected 435 tasks through direct execution and the real interactive terminal. It originally flagged 274 tasks; source review removes two permission-check false positives and one unsupported account-existence assertion, leaving 271 phase 3 exclusions. The remaining 164 review outcomes mean no actionable mismatch in the implemented assertions. Source coverage is incomplete and no Docker reference run was made.
| First exclusion | Tasks | Reason |
|---|---|---|
| Phase 3 | 271 | Source-declared user identity, ownership or permissions differ under Apptainer. The categories overlap; these tasks are counted once. |
| Phase 4 | 53 | The reference solution returns reward 0, including the restored ICMP-dependent task. |
| Phase 5 | 22 | Tests accept the task without an answer and return reward 1. |
| Phase 5 | 1 | The verifier repeatedly writes reward 1 without executing tests; this is an invalid no-op result. |
The two corrected permission cases used an incorrect path after a source cd, or an obsolete mode expectation overwritten by a later recursive chmod. The account case inferred guaranteed creation from useradd ... || true; it remains a review outcome. Original observations and the review are included in the evidence.
Known retained defect: task_000000_9719b9f3 passes gold and no-op but its reference solution can supply a fake YAML module accepted by a version-only verifier. It is restored because phase 1 filtering is disabled. Phase 3–5 acceptance does not resolve this reference-integrity defect.
Repairs
The task patcher makes Docker setup usable under Apptainer while preserving the source shell environment, working directory and ordered initialization. It shares images only for matching dependency recipes, moves initialization into guarded per-task setup, and provides offline verification with the source assertions preserved. Five truncated prompts recover their complete source descriptions.
The existing reference-solution repairs remain: pinned pip installations for eight tasks, CPU-count repairs for five tasks, and a verifier subprocess PATH repair for one task. Removing the phase 1 instruction addition changes no environment, setup, solution or verifier files. The complete upstream comparison and reproduction scripts are included below.
Teacher results: pass@k
Qwen/Qwen3-30B-A3B-Instruct-2507 used Terminus 2, with four attempts on each of 2,141 tasks: 8,564 completed attempts. These are historical measurements from before removal of the instruction addition. Four restored publication tasks have no teacher measurements; rates below cover the measured 2,141 tasks, not all 2,145 retained tasks. Teacher outcomes do not exclude tasks.
| Model | Measured tasks | Attempts per task | Pass@1 | Pass@2 | Pass@3 | Pass@4 |
|---|---|---|---|---|---|---|
| Qwen3-30B-A3B-Instruct-2507 | 2,141 | 4 | 42.73% | 53.09% | 58.43% | 61.79% |
1,323 / 2,141 tasks were solved at least once; 3,659 / 8,564 attempts succeeded. Pass@k uses the unbiased estimator from each task's four attempts. The saved results include 846 agent timeouts, 4 terminal-protocol errors, and 1 test-upload error. Five attempts have no raw verifier reward and count as unsolved under the recorded terminal-error policy; all 8,564 trajectories and result records are present.
Among the original 161 conversion-review cases, 155 entered teacher generation after gold/no-op filtering. They completed 620 attempts; 101 / 155 were solved, with 46.13% pass@1 and 65.16% pass@4. No recorded container, terminal-protocol or test-upload exceptions occurred in that cohort. This is useful runtime evidence, not complete conversion certification. One HTTP case has an unresolved possible shared-port collision and needs an isolated repeat to establish its cause.
Runtime and images
Validation used Apptainer on Helma. Teacher jobs used one H200 and 32 CPUs per job, concurrency 10, a 65,536-token context and 16,384-token reply limit, with temperature 0.7, top-p 0.8 and top-k 20. All 15 teacher jobs completed in approximately 55–79 minutes.
This PR includes 1,835 pristine native SIF bundles covering all 2,145 retained tasks, totaling approximately 210.5 GiB. The image release manifest binds each task's environment context to its immutable bundle and SHA-256 checksums. Files use the publisher's images/bundles-v1/<bundle-id>/<image-name>/... format; Docker archives and live trial containers are not published. Earlier successful builds retain their recorded base provenance; later retry builds used the pinned local base. The image manifest pins the immutable image commit.
Files and evidence
The files below are included in this PR.
- Datasource README, including pass@1–4.
- Retained tasks and archived tasks with reasons.
- Run record and per-task findings.
- Validation reports: phase 3, phase 4, phase 5.
- Reviewed conversion assertions and original conversion observations.
- Teacher pass@1–4, per-task historical pass counts, phase 6, and phase 7.
- Teacher trajectories and result records, trajectory manifest, and conversion-case review.
- Upstream patch comparison, task patcher, dry-run preparation script, input provenance, and historical input binding.
- Image release manifest.