FORGE-Bench: Public Test Split

This repository contains the 1,000 public test inputs for FORGE: Fault-Oriented Rework Code Generation for Executable Interlocking Assembly. The intended Hugging Face repository is QLU-LLM/FORGE-Bench.

The paper describes a benchmark of 4,000 tasks across 26 assemblies, with 3,000 training tasks across 21 assemblies and 1,000 test tasks across 5 held-out assemblies. Only the test split is distributed in this repository. The paper's full benchmark size must not be confused with the number of records available here.

Files

README.md
.gitattributes
test.jsonl
data_types.py
SHA256SUMS.txt
metadata/
  dataset_manifest.json
  assembly_manifest.json

Released split	Records	Assemblies	Reference answers
test	1,000	5	Not included

No training or validation records are included. No model predictions, experiment results, human assessment records, LLM judge records, or scoring scripts are included. This release uses JSONL; no Parquet file is supplied.

Record format

Each line in test.jsonl is one JSON object:

Field	Type	Meaning
id	string	Original sample identifier
assembly_id	string	Assembly identifier
split	string	Always test
public_task	object	Structured task input

public_task contains task, initial_state, target_state, actual_state, assembly_graph, blocking_graph, contact_graph, constraint_graph, action_space, part_catalog, and task_public_metadata.

The supplied action space declares remove, insert, reset_pose, replace, rotate_translate_to_unlock, and translate_rotate_to_lock. The JSONL file is a byte-for-byte copy of the existing public test release: record IDs, ordering, and task contents have not been changed for this organization release. Some identifiers retain historical version strings; they are preserved for traceability.

Assembly distribution

Assembly	Test records
parametric_helical_gate_11	260
parametric_offset_core_09	200
parametric_pyramid_lock_05	170
parametric_t_lock_07	170
parametric_tower_lock_10	200
Total	1,000

Loading

Load locally using the Python standard library, preserving the nested task objects:

import json

with open("test.jsonl", encoding="utf-8") as handle:
    test = [json.loads(line) for line in handle if line.strip()]

assert len(test) == 1000
print(test[0]["id"])

After publication, download the original JSONL from the Hub:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="QLU-LLM/FORGE-Bench",
    filename="test.jsonl",
    repo_type="dataset",
)
with open(path, encoding="utf-8") as handle:
    test = [json.loads(line) for line in handle if line.strip()]

data_types.py also provides a small standard-library JSONL reader.

Scope and availability

This is a structured-input dataset release. Reference annotations and reference rework plans are not included, so this repository alone does not provide ground-truth scoring. The dataset concerns symbolic assembly states and action constraints; it does not provide a complete continuous-physics simulator.

Some input fields retain asset references such as assets/*.stl. The referenced STL/FBX assets are not bundled in this repository. References such as constraint_graph.json identify components of the original task format; the corresponding graph object is embedded in each record's public_task.

License

CC BY-NC 4.0, consistent with the previously selected dataset license. See https://creativecommons.org/licenses/by-nc/4.0/ .