Datasets:
Add CalibForge stage-3 and stage-3-plus-5 RL ablations
CalibForge RL ablations: 5,377 tasks after build/setup validation, and 5,320 after additionally requiring valid no-op reward 0. The second variant holds out 57 tasks: 9 archived cases and 48 unresolved verifier results.
| Variant | Tasks | Stage 3 | Stage 5 |
|---|---|---|---|
calibforge/ablations/stage3/tasks.parquet |
5,377 | Passed | Not an acceptance gate |
calibforge/ablations/stage3_stage5/tasks.parquet |
5,320 | Passed | Passed |
Includes the full validation run and reruns of repaired tasks. Thirteen verifiers now report missing required agent outputs, a required dependency, or unfinished requested service/build work as ordinary failed tests with reward 0 instead of aborting collection/setup. Original behavioral checks remain when their prerequisites are present. The latest ten repairs cover eight agent-created output files, PostgreSQL recovery, and the Go/cgo service build; all ten passed fresh stage-3 preparation and stage-5 no-op validation. These thirteen payload differences are recorded in variants.json, so this comparison is not a pure selection-only ablation.
The nine archived stage-5 cases comprise four no-op reward failures, one memory-exhausting verifier, and four upstream images missing Python despite requiring Python/pytest for grading. The remaining 48 cases are held out, not declared unsolvable:
| Unresolved finding | Tasks |
|---|---|
| Dependency/interpreter/pytest failures | 21 |
| Native-code crashes in tasks asking the agent to repair that code | 8 |
| Memory-limit failures | 7 |
| Verifier timeouts | 7 |
| Backend permission restrictions | 2 |
| Shell harness failures | 2 |
| Unexpected missing baseline file | 1 |
Permission restrictions are specific to this backend: writing /etc/hosts and binding port 80. Resource failures and native crashes do not establish universal task unsolvability. The original 54 build/backend exclusions precede both variants.
Stage-1 checking, with the corrected Dockerfile checks and audited CPU-resource exceptions, yields 4,498 passes out of 5,372 tasks. The stage-1 selection excludes 871 tasks with unresolved pip pins, one task with pinned apt packages, and two tasks with unrecognized configuration fields. Non-failing path warnings remain diagnostic. The dependency-preparation comment contains no test contents or solutions and no longer counts as an executable reference; unrelated version assignments no longer count as apt pins. Five measured nproc uses respect their Helma Slurm CPU allocations. The task-specific exclusion list and corrected results are included in the evidence. The combined stage-1/3/5 Parquet remains pending; the two published ablations above retain their existing selections. Rubric-filtered variants will be tracked separately.
Every Parquet row includes stages_passed (3 or 3,5) and validation_variant (stage3 or stage3_stage5). Each variant has a selection manifest and payload hashes. Supporting outcomes, verifier-repair provenance, original and current unresolved-case lists, and the investigation are included under calibforge/ablations/evidence/. These are staged training ablations, not a claim of a fully validated final CalibForge release.