Datasets:

ComerX
/

forgebench-v1

license: mit
language:
  - en
task_categories:
  - question-answering
  - text-generation
size_categories:
  - n<1K
configs:
  - config_name: forgebench_full
    data_files:
      - split: tasks
        path: data/forgebench_full.jsonl
  - config_name: schema_example
    data_files:
      - split: examples
        path: data/schema_example.jsonl
tags:
  - minecraft
  - neoforge
  - forge
  - mixin
  - code
  - benchmark
pretty_name: ForgeBench v1

Check the newer (v4) version

ForgeBench v1 — independent benchmark for Minecraft-modding LLMs

Contamination-free evaluation benchmark for models fine-tuned on the ComerX/forge-neoforge-sft-* dataset family. Targets NeoForge 1.21.1 and Forge 1.20.1 (Mojang mappings). 100 tasks, 5 modules × 20.

Data files (what the Dataset Viewer loads)

Config	Split	File	Rows
`forgebench_full`	`tasks`	`data/forgebench_full.jsonl`	100
`schema_example`	`examples`	`data/schema_example.jsonl`	3

Both files are UTF-8 JSONL, one task object per line — loadable with load_dataset("ComerX/forgebench-v1", "forgebench_full") or line-by-line with json.loads. Everything else in the repo (harness/, template/, reports) is tooling, not data.

Task schema (strict)

{
  "task_id": "MIXIN_014",
  "module": "api | mixin | migration | debug | build",
  "instruction": "prompt given to the evaluated LLM",
  "context": "vanilla snippet / crash log / legacy code (may be empty)",
  "golden_reference": "reference answer",
  "eval_criteria": {
    "must_compile": true,
    "required_annotations": ["@Inject"],
    "forbidden_apis": ["net.minecraft.util.text"],
    "gametest_required": false
  }
}

context is "" for tasks that need no reference material (e.g. pure API registration drills); it contains vanilla/legacy code or FML log excerpts for mixin, migration and debug tasks.

Modules

Module	Count	Task type	Grading
`api` (API & registration)	20	instruction→code	compile + AST gate
`mixin` (Mojmap targets)	20	FIM / translation	compile + AST gate
`migration` (1.20.1→1.21.1)	20	refactoring	compile + AST gate (legacy APIs forbidden)
`debug` (FML crash logs)	20	log analysis → JSON	judge + JSON-shape gate
`build` (Gradle KTS / mods.toml)	20	config generation	judge + required-token gate

Layout

data/
  forgebench_full.jsonl    # 100 tasks (single JSONL, one task per line)
  schema_example.jsonl     # 3 complete examples (Mixin / Migration / Crash log)
harness/
  Dockerfile               # sandbox image: JDK 21 + Gradle 8.9 + prebuilt MDK template
  sandbox_runner.py        # injects answer into MDK -> ./gradlew build -x test -> Pass@1
  ast_eval.py              # javalang AST gate: syntax, forbidden APIs, hallucinated
                           # imports, required annotations, version-mix penalties
  gametest_runner.py       # GameTest template generation + headless run
  judge_prompt.txt         # strict LLM-as-a-judge rubric for debug/build tasks
  run_benchmark.py         # driver: aggregates sandbox + AST + judge into a report
  requirements.txt
template/                  # minimal NeoForge 1.21.1 MDK (NeoGradle 7 userdev)
VERIFICATION.md            # benchmark self-verification results (19/19)
verification_result.json   # machine-readable check log
SHA256SUMS.txt             # integrity manifest (sha256 of every repo file)

Running

# 0. build the sandbox image (once)
docker build -t forgebench-sandbox:1.21.1 .

# 1. produce answers.jsonl: {"task_id": ..., "answer": ...} per row

# 2a. offline static gate only
python3 harness/run_benchmark.py --data data/forgebench_full.jsonl \
    --answers answers.jsonl --skip-compile --out report.json

# 2b. full compile evaluation (inside Docker)
python3 harness/run_benchmark.py --data data/forgebench_full.jsonl \
    --answers answers.jsonl --out report.json

# 3. judge-graded tasks (debug/build): run your judge LLM with harness/judge_prompt.txt
#    and pass scores via --scores '{"DEBUG_004": {"score": 9, "verdict": "pass"}, ...}'

Contamination status

Golden references and instructions were screened against the public SFT corpus (ComerX/forge-neoforge-sft-v2.1.0-rc2 + -rc1, sampled shards): 0 hits (normalized 300-char prefix containment). The benchmark tasks are synthetic examples, not derived from any corpus repository.

Integrity

Frozen artifact: SHA256SUMS.txt (sha256 of every file in this repo).
The benchmark is immutable once published; changes require a new version tag.

License

MIT — applies to the benchmark data (data/), the harness (harness/), and the MDK template (template/).