forgebench-v2 / VERIFICATION.md
ComerX's picture
ForgeBench v2: dual-target (forge-1.20.1 / neoforge-1.21.1), 130 tasks, porting module, verified harness
91485a4 verified
|
Raw History Blame Contribute Delete
2.56 kB

ForgeBench v2 — self-verification report (publication-time)

Date: 2026-10-27 · Host: constrained sandbox (PolarFS workdir, overlay upper-layer ~full)

1. Schema validation — PASS

  • 130/130 rows contain all 9 required fields; eval_criteria contains all 5 keys.
  • task_id uniqueness: OK. target/task_scope enum values: OK.
  • Cross-check: forge-1.20.1 goldens contain no import net.neoforged*; neoforge-1.21.1 goldens contain no import net.minecraftforge* (comment mentions allowed).
  • multi_file rows: context and golden_reference are dicts with project-relative paths.

2. Golden self-test (ast_eval --self-test semantics, in-process) — PASS 130/130

Every golden_reference evaluated against its own target profile:

  • 130 passed, 0 failed, 0 penalty warnings.
  • Harness upgrades applied during verification (part of the v2 harness): a) tokenizer fallback: sources that tokenize but exceed javalang's Java-8 grammar (records/switch expressions) are analyzed via regex imports/annotations with a PARSE_LIMITATION info finding (no penalty; sandbox compile stays authoritative); b) non-java resource gate (json/toml/yaml/gradle) now strips // and # comment lines before forbidden-API matching (fixes false positives on migration comments); c) .yml/.yaml added to resource suffixes (GitHub Actions golden in BUILD_017).

3. Contamination check vs SFT corpus — CLEAN

  • Baseline: local package of ComerX/forge-neoforge-sft-v2.1.0-rc1 (8,357 records, 45 jsonl files, 15 families).
  • Method: normalized 8-line-window exact-match between all v2 golden files and every SFT record text (messages + context_files), comments/blank lines dropped.
  • Raw result (k=8): 4 window hits on 2 records — manual inspection shows BOTH are idiomatic boilerplate collisions (ModDevGradle runs{} block; standard AbstractContainerScreen.render override), not copied content.
  • Strict result (k=12): 0 hits → verdict CLEAN.
  • Machine-readable: reports/contamination_result.json.

4. Docker compile gate — NOT EXECUTED (infra limitation, disclosed)

  • Publication host: no guaranteed Docker daemon; memory ceiling ~4 GB < NeoForm ≥8 GB.
  • Status: harness shipped as designed-not-executed. Re-verify on a ≥8 GB runner before quoting Pass@1 numbers. Both template targets carry a MEMORY CONTRACT header.

5. Artifacts

  • data/forgebench_v2_full.jsonl (130 rows) — SHA256 in SHA256SUMS.txt
  • verification scripts: scripts/ in the publication workdir (merge_v2.py, verify_goldens.py, contamination_check.py)