Upload BUILD.md with huggingface_hub

BUILD.md

@@ -0,0 +1,78 @@
+# dataset_v2 — CodeNet Java compile errors, source-name aware
+
+Status: **compiled corpus, unlabelled**. Published at
+<https://huggingface.co/datasets/criyle/codenet-compile-errors-v2> (private
+false). Do not mix it with `dataset_v1/`: v1 uses the old `Main.java` convention
+and the 65-leaf v1 taxonomy; v2 has no labels yet.
+
+## Build
+
+| Step | Command / source |
+|---|---|
+| Source | `iNeil77/CodeNet`, Java split |
+| Filter | `status == "Compile Error"` (53,023 submissions) |
+| Sandbox | `http://10.0.0.162:5050`, OpenJDK 21.0.12.1 |
+| Compile | javac under the declared public top-level class name; `Main.java` fallback |
+| Build | `uploader/04_build_dataset_v2.py` |
+| Publish | `uploader/05_publish_v2.py --repo criyle/codenet-compile-errors-v2` |
+
+## Result
+
+| Outcome | Rows |
+|---|---:|
+| Ingested `Compile Error` submissions | 53,023 |
+| Compiled cleanly once named after the class (file-name artefacts) | **18,465** |
+| Still failing (the corpus) | **34,555** |
+| javac internal crashes (excluded) | 3 |
+| Rows with no `error:` line | **0** |
+
+The source-name fix is the main difference from v1: `s053550749`, for example,
+was a `Class name mismatch with file name` row under `Main.java`; it compiles
+cleanly as `MyClass.java` and is excluded here. Of the 12,208 v1 rows labelled
+`Class name mismatch with file name`, **10,741 now compile cleanly** and **944
+still fail** under the declared name (132 of the remaining corpus rows are
+multiple-public-class errors, where no single file name satisfies javac).
+
+## Corpus columns
+
+`row_id, sid, submission_id, pid, problem_id, uid, user_id, date, date_iso,
+source_name, code, compilemessage, compile_status, exit_code, javac_version,
+category, label_status, taxonomy_version, session_id, session_6h, source,
+original_status, code_size`
+
+- `category` is empty and `label_status=unlabelled`; `taxonomy_version=2.1`.
+- `session_id` is a `(user, problem)` 2-hour session, computed over the **full**
+  Java split (accepted submissions included); `session_6h` is the 6-hour variant.
+- 7,251 users, 1,747 problems, **20,567 sessions** (mean 1.68, max 28 rows),
+  100% session coverage.
+- Date range 2010-01-04 → 2020-10-01.
+
+## Harness bug found and fixed during the build
+
+The stock `go_judge` client asks the sandbox to copy out `Main.class`. When a
+clean submission declares only package-private classes (e.g. `class q2`) it
+produces `q2.class`, and go-judge returns
+`File Error: open Main.class: no such file` even though javac exited 0. The first
+pass therefore classified **3,583 clean submissions as `compile_error`**. A
+compile-only request with no `copyOutCached` rechecked those rows: **3,580 were
+clean, 3 were real errors**. The final corpus has 0 rows without an `error:` line.
+The three `An exception has occurred in the compiler (21.0.12.1)` rows are
+excluded as `compiler_crash`.
+
+## Files
+
+- `corpus_v2.parquet` — 34,555 still-failing submissions (the published train split)
+- `compiled_all.parquet` — 53,023 compile attempts, including artefacts and crashes
+- `session_map_all.parquet` — sessions over the full Java split
+- `source_manifest.json` — counts and SHA-256 hashes
+- `categories.v2.yaml` — active v2.1 taxonomy
+- `codenet_v2.db` — resumable build DB (gitignored)
+
+## Reproduce
+
+```bash
+.venv/bin/python uploader/04_build_dataset_v2.py
+.venv/bin/python uploader/04_build_dataset_v2.py --recheck-no-error
+.venv/bin/python uploader/04_build_dataset_v2.py --export-only
+.venv/bin/python uploader/05_publish_v2.py --dry-run
+```