Document CrossCodeEval release statistics, rules and recorded runtime

crosscodeeval_environment-reliable/README.md

@@ -1,43 +1,47 @@
 # CrossCodeEval
 
-The pinned TaskTrove release contains **7,763 cross-file code-completion tasks** across C#, Java, Python and TypeScript. All existing context, scoring and instruction repairs are preserved.
+## Dataset stats
 
-- **Docker images: 3 distinct definitions originally; 3 retained, provided as 3 shared cached images.** No image consolidation was needed.
-- **Tasks: 6,710 kept, all 6,710 changed; 0 unchanged and 1,053 excluded.**
+- **Tasks:** 7,763 original → 6,710 retained, all 6,710 patched; 1,053 archived.
+- **Images:** 3 original → 3 current definitions; 3 [cached Apptainer bundles](https://huggingface.co/datasets/FWeindel/validated-tasks/tree/main/crosscodeeval_environment-reliable/cached_images). No consolidation.
 
-| Language | Original | Kept and changed | Excluded |
-| --- | ---: | ---: | ---: |
-| C# | 1,768 | 1,353 | 415 |
-| Java | 2,139 | 1,895 | 244 |
-| Python | 500 | 432 | 68 |
-| TypeScript | 3,356 | 3,030 | 326 |
+## Dataset characteristics
 
-## Stage 3: Build/setup
+Complete the next statement in a real C#, Java, Python or TypeScript repository using excerpts from related files.
 
-- **6,710 tasks patched:** pinned `pytest==9.1.1` and `pytest-timeout==2.4.0` in their image definitions for reproducible installation.
-- **0 tasks excluded.** All 6,710 environments built and started successfully.
+<details><summary>Domains</summary>
 
-## Stage 4: Reference solution reward 1
+**Software Engineering:** cross-file code completion, repository APIs and library usage.
 
-- **6,710 tasks patched:** added a reference solution that writes the benchmark's expected completion, allowing the correct answer to be checked automatically.
-- **0 tasks excluded.** All 6,710 reference solutions scored 1.
+</details>
+<details><summary>Capabilities</summary>
 
-## Stage 5: No-op reward 0
+**Exploration:** find required names and calling conventions in supplied context. **Specification Adherence:** write the requested continuation in the required output file.
 
-- **6,710 tasks patched:** replaced the original grading scripts with a scorer that compares the completed statement after normalizing comments and whitespace. This removes rewards for incomplete prefix matches and avoids rejecting correct Java completions solely for indentation.
-- **0 tasks excluded.** All 6,710 no-op attempts scored 0.
+</details>
 
-The stage repair counts overlap; they refer to the same 6,710 retained tasks.
+**Intended benchmark transfer:** Python subset: HumanEval+, MBPP+ and DS-1000 mini-200. Java/Python/TypeScript: repository exploration for SWE-bench Verified random-100. Labels follow [CLI-Universe](https://arxiv.org/abs/2606.22883v1).
 
-## Task-quality changes retained
+## Per-stage breakdown
 
-- **6,710 tasks:** restored the benchmark's retrieved excerpts from related files, excluding excerpts from the target file or containing the answer.
-- **6,710 tasks:** preserved the corrected instructions, context-file guidance and automatic timeout/anti-cheating sentence.
-- **1,043 tasks excluded:** required names could not be inferred from the prompt, retrieved context or known library names.
-- **10 tasks excluded:** the scored completion contained no variable, function, class or other code name.
+Counts cover retained tasks; repair categories overlap.
 
-## Images and validation
+- **1, rule-based checks:** 6,710 retain the timeout/anti-cheating instruction (`instruction-suffix`) and context guidance; 0 archived by this validation stage. Patcher rules separately archived 1,043 tasks with required identifiers absent from available context and 10 whose scored answer contained no identifiers. All 6,710 retain restored cross-file excerpts with answer leakage removed.
+- **3, build/setup:** 6,710 received pinned verifier dependencies; 6,710 passed, 0 archived. Docker definitions were converted to Apptainer for the Marin Harbor bridge and repository runtime patches.
+- **4, reference reward 1:** reference solutions added to all 6,710; all passed, 0 archived.
+- **5, no-op reward 0:** scoring repaired in all 6,710 to reject incomplete prefix matches and normalize comments/whitespace; all passed, 0 archived.
 
-Download the three cached Apptainer images from [images/](https://huggingface.co/datasets/FWeindel/validated-tasks/tree/main/crosscodeeval_environment-reliable/cached_images). The Parquet includes pinned image references for automatic download and checksum verification.
+Results are from the [historical validation](https://huggingface.co/datasets/FWeindel/validated-tasks/blob/b8a56dc21aa9f78340048a367e70852c59d01ff8/runs/2026-10-03-c536bae24db8.json); retained task payloads are unchanged.
 
-Task payloads are unchanged from the [previous release](https://huggingface.co/datasets/FWeindel/validated-tasks/discussions/14); its recorded stages 1, 3, 4 and 5 remain applicable. [Teacher results](https://huggingface.co/datasets/FWeindel/validated-tasks/discussions/6) are available separately.
+## Resources and timing
+
+Helma, x86_64, **48 CPUs / 128G**, no GPUs, concurrency 16; Python **3.12.14**, Harbor **0.8.1**, Apptainer **1.5.4-1.el9**, host networking through the cluster proxy. Marin Harbor revision [`7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d).
+
+| Historical validation stage | Wall time | Mean per task (6,710 tasks) |
+| --- | ---: | ---: |
+| Static checks | 1,528.5 s | 3.36 s |
+| Build/setup | 6,938.0 s | 16.50 s |
+| Reference | 6,061.3 s | 14.39 s |
+| No-op | 6,468.8 s | 15.35 s |
+
+Means cover summed check execution, environment start/inspection/stop, and one reference/no-op trial per task, respectively. Image-build and patching time: **Not recorded**.