Keep dataset naming guidance concise

README.md

@@ -26,43 +26,16 @@ Datasources cleaned and filtered using the [OT-next-data validation protocol](ht
 
 ## Dataset versions
 
-Folders follow `<dataset_name>_<filtering-level>`, for example
-`camel-ai__SETA-Env_environment-reliable`. Preserve the data-source identifier;
-`__` separates an upstream owner and dataset name when applicable.
+Folders follow `<dataset_name>_<stage>`. Each stage builds on the previous one.
 
-The three filtering levels are cumulative: each later version adds to the previous one.
-
-| Filtering level | Protocol stage | Purpose |
-| --- | --- | --- |
-| `environment-reliable` | 1 | Environments build and run reliably and reproducibly, with expected reference and no-op rewards where those checks are available. |
-| `task-quality` | 2 | Adds rule-based checks and LLM task review for instruction/verifier alignment, hidden requirements and solution or test leakage. |
-| `trajectory-quality` | 3 | Adds filtering informed by teacher trajectories, trace heuristics, LLM trajectory review and pass@k. |
-
-These are filtering levels, not numbered releases. Pin a Hugging Face commit
-with `revision=` for a reproducible snapshot. Consult the recorded validation
-evidence for exact check coverage, skipped checks, sampling and thresholds;
-the folder name alone does not certify every check. Merely collecting teacher
-trajectories does not make a dataset `trajectory-quality` unless those results
-were used for filtering.
-
-Currently published:
-
-- `camel-ai__SETA-Env_environment-reliable`: SETA, with build, reference and no-op validation.
-- `Calibforge_environment-reliable`: CalibForge, with build and no-op validation.
-  Reference validation is unavailable because the source provides no reference solutions.
-
-The `task-quality` and `trajectory-quality` names define subsequent filtering
-levels; those folders are not yet published on `main`. Renaming the existing
-folders does not rerun validation or add checks to their recorded coverage.
-
-## Files
+| Stage suffix | Filtering |
+| --- | --- |
+| `environment-reliable` | Ensure environments build and run reliably, reproducibly and robustly, and produce the expected rewards for the reference solution and no-op agent. |
+| `task-quality` | Apply rule-based checks and LLM-as-a-judge review to catch instruction/verifier misalignment, hidden requirements and solution or test leakage. |
+| `trajectory-quality` | Use teacher trajectories for further filtering through LLM-as-a-judge review, heuristic checks on traces and pass@k-based filtering. |
 
-- `tasks.parquet`: retained tasks.
-- `archive.parquet`: excluded tasks and their recorded reasons.
-- `README.md`: data-source description and usage.
-- `patch.py`: the published patcher, when included.
-- `cached_images/`: cached image bundles and checksums, when included.
-- `pass_counts.parquet`: per-task teacher success counts, when published.
+CalibForge has no reference solutions, so reference validation is skipped.
+`tasks.parquet` contains retained tasks; `archive.parquet` contains exclusions.
 
 ## Download retained tasks and images
 
@@ -83,13 +56,5 @@ This also downloads the cached build images for the tasks. Existing image
 references in the task payloads remain pinned to their original commits, where
 the referenced paths remain available.
 
-## Previous folder names
-
-| Previous name | Current name |
-| --- | --- |
-| `camel-ai__SETA-Env_filtered_stages_3_4_5` | `camel-ai__SETA-Env_environment-reliable` |
-| `Calibforge_filtered_stages_3_5` | `Calibforge_environment-reliable` |
-
-The old suffixes described runner stage IDs. Runner IDs in commands and reports
-remain unchanged; the protocol now groups them into three filtering stages.
-Historical commits retain the previous paths.
+Previous `_filtered_stages_<numbers>` folder names are now renamed to
+`_environment-reliable`. Historical commits retain the old paths.