--- language: - en task_categories: - text-generation tags: - agent - code - agentic-tasks - harbor - reinforcement-learning configs: - config_name: default data_files: - split: train path: '*_v1/tasks.parquet' - config_name: v2 data_files: - split: train path: '*_v2/tasks.parquet' - config_name: archive data_files: - split: train path: '*_v1/archive.parquet' - config_name: calibforge_v4 data_files: - split: train path: calibforge_v4/tasks.parquet - config_name: calibforge_v4_judgments data_files: - split: train path: calibforge_v4/results.parquet - config_name: calibforge_v4_judge_attempts data_files: - split: train path: calibforge_v4/judge_attempts.parquet --- # Validated tasks Datasets repaired and filtered using the [OT-next-data validation protocol](https://github.com/MLI-lab/OT-next-data/blob/main/validation/PROTOCOL.md). - **V1:** environment reliability and rule-based task-quality repairs. Build/setup, no-op and reference checks where reference solutions exist. - **V3:** newly solved tasks absent from v2, with the new evaluation's traces and statistics. Combine v2 and v3 for the full solved set. - **V2:** keep tasks solved at least once by an evaluated model. Includes partial results where a success is already recorded; each README lists models, attempt counts and coverage. ## Planned layout We plan to reorganize this repository so each dataset stores its tasks once. Results from every validation stage and evaluation will go in one table instead of separate v1, v2 and v3 folders. This is not in place yet; the folders below still describe the current layout. - **`results.parquet`:** one row per task revision and one column per check, with the value `pass`, `fail`, `warning`, `not_run` or `not_recorded`. Some checks add a details column, for example the model and settings for pass@k or the failed criteria for an LLM review. New checks and evaluations add new columns. - **`tasks/`:** `r1.parquet` holds all original tasks. Later revision files hold only the tasks that changed. - **Changed tasks:** a changed task gets a new row with a new `content_sha256` and a `change_reason`, and its checks start as `not_run` until rerun. Filter on `latest = true` for the current dataset. ## Incremental v3 releases V3 task files contain only newly solved tasks, with no task-ID overlap with the corresponding v2 task file. Historical traces remain in v2; v3 includes every attempt from its new evaluation, successful and failed. These subsets were harder for the earlier evaluated models, not necessarily for every future model. | Dataset | v2 tasks, available separately | Newly solved v3 tasks | Combined unique tasks | | --- | ---: | ---: | ---: | | [BugsInPy](bugsinpy_v3/README.md) | 65 | 406 | 471 | | [CalibForge](calibforge_v3/README.md) | 1,512 | 2,287 | 3,799 | | [EndlessTerminals](endless_v3/README.md) | 1,958 | 121 | 2,079 | | [NL2Bash](nl2bash_v3/README.md) | 938 | 104 | 1,042 | | [TermiGen](termigen_v3/README.md) | 2,277 | 230 | 2,507 | Each release README includes a v2 + v3 loading example. Counts here describe the incremental releases published on 2026-10-26. Earlier cumulative v3 versions remain accessible by their pinned HF commit revisions. ## Earlier v1/v2 size snapshot **V1: 42,166 kept, patched tasks across 8 data sources.** **V2: 7,684 tasks solved at least once across the same 8 data sources.** V2 is a subset of V1, not additional tasks. The table below is an earlier snapshot and predates subsequent evaluations. For current counts, use the individual release READMEs and the v3 table above. Percentages refer to this earlier snapshot. | Dataset | V1 tasks | Share of V1 | V2 tasks | Share of V2 | | --- | ---: | ---: | ---: | ---: | | BugsInPy | [489](bugsinpy_v1) | 1.16% | [65](bugsinpy_v2) | 0.85% | | CrossCodeEval | [6,710](crosscodeeval_v1) | 15.91% | [863](crosscodeeval_v2) | 11.23% | | TermiGen | [3,336](termigen_v1) | 7.91% | [2,277](termigen_v2) | 29.63% | | SETA | [3,107](seta_v1) | 7.37% | [2,205](seta_v2) | 28.70% | | CalibForge | [5,317](calibforge_v1) | 12.61% | [84](calibforge_v2) | 1.09% | | ScaleSWE | [16,786](scaleswe_v1) | 39.81% | [292](scaleswe_v2) | 3.80% | | SWE-Lego | [4,277](swelego_v1) | 10.14% | [575](swelego_v2) | 7.48% | | EndlessTerminals | [2,144](endless_v1) | 5.08% | [1,323](endless_v2) | 17.22% | | **Total** | **42,166** | **100%** | **7,684** | **100%** | `tasks.parquet` contains kept tasks, `archive.parquet` archived ones. ScaleSWE still uses its original Docker Hub images; we haven't consolidated them into shared images yet. ## Download tasks and images ```python from huggingface_hub import snapshot_download snapshot_download( "FWeindel/validated-tasks", repo_type="dataset", allow_patterns=["seta_v2/tasks.parquet", "seta_v1/cached_images/**"], local_dir="data", ) ``` ## Trace stats Trace stats put each attempt into one outcome category: | Outcome | Meaning | | --- | --- | | Task complete | The agent said it was done and confirmed that it wanted to finish. The verifier still checks whether the solution is correct. | | tmux session closed | The terminal closed before normal completion, for example because the agent sent an `exit` command or killed tmux. | | Task timeout | The agent ran out of time. | | Repeated output truncation | Three responses in a row were too long (current default: 8,192 output tokens) and could not be used. | | Other limits | The agent hit another limit, such as the number of turns, response length, context size or task memory. | | Environment error | The container could not start, or a terminal command or communication failed. Confirmed terminal closures go under “tmux session closed.” | | Verifier error | The checker could not run or finish normally. A solution simply failing its tests is not a verifier error. | | Model server error | A request to the model failed, for example because the server was unavailable or the request timed out. | | Harness error | The software running the agent failed, for example because of a bug in retry handling. | | Other error | An error was recorded, but it does not fit the categories above. | | Not recorded | The saved result does not say how the attempt ended. | These labels come from the harness's recorded stop reason and error type, not an LLM judge. ## CalibForge v4 judge annotations [calibforge_v4](calibforge_v4/README.md) keeps the full reviewed v1 task pool and adds judge annotations without filtering tasks.