Zenng2812/arca-synthetic-adaptive

ARCA Synthetic Adaptive Graph-Path Constraints

This dataset is a controlled synthetic benchmark for testing whether Adaptive Residual Constrained Attention (ARCA) reacts to external constraint quality.

It does not use LLMs, pretrained language models, or natural-language generation. Every sample is generated by Python with exact ground truth.

Task

Each base sample contains a random directed graph. Nodes have discrete values. A query gives:

• a start node
• a relation path such as r2 -> r1 -> r4

The target is the value of the final node reached by following the relation path in the original graph.

For the same input graph/query/target, the dataset emits several constraint variants. The input graph remains unchanged; only constraint_edges changes.

This is designed to isolate the ARCA question:

If the factual input is still present but the external attention constraint becomes incomplete or wrong, does ARCA's residual path and learned beta adapt?

Splits

Split	Rows	Reasoning depth	Nodes
train	10,000	2-4 hops	8-12
validation	2,000	2-4 hops	8-12
test_iid	2,000	2-4 hops	8-12
test_ood	2,000	5-6 hops	15-20

Rows are constraint-variant examples. Most base graph inputs appear with all six constraint variants and share the same base_sample_id.

Constraint Variants

Constraint type	Meaning
perfect	Constraint edges equal the original graph edges
random_delete	Deletes non-gold constraint edges while preserving the gold path
critical_delete	Deletes at least one gold-path edge from the constraint
severe_critical_delete	Deletes multiple gold-path edges when possible
false_edge_addition	Adds false/distractor edges to the constraint while keeping the gold path
mixed_corruption	Deletes gold and non-gold edges and adds false edges

Important: when a gold edge is deleted from constraint_edges, the corresponding fact still remains in edges. Only the external attention constraint is corrupted.

Schema

Each row is a JSON object:

{
  "sample_id": "train_0000000_6cb21b259e1f764e36fc22dc_perfect",
  "base_sample_id": "train_0000000_6cb21b259e1f764e36fc22dc",
  "split": "train",
  "nodes": ["N_AAAAA", "N_BBBBB"],
  "node_values": {"N_AAAAA": 4, "N_BBBBB": 7},
  "edges": [{"src": "N_AAAAA", "relation": "r2", "dst": "N_BBBBB"}],
  "query_start": "N_AAAAA",
  "query_relations": ["r2"],
  "target_node": "N_BBBBB",
  "target_value": 7,
  "gold_path_nodes": ["N_AAAAA", "N_BBBBB"],
  "gold_path_edges": [{"src": "N_AAAAA", "relation": "r2", "dst": "N_BBBBB"}],
  "constraint_edges": [{"src": "N_AAAAA", "relation": "r2", "dst": "N_BBBBB"}],
  "constraint_type": "perfect",
  "corruption_rate": 0.0,
  "num_nodes": 2,
  "num_edges": 1,
  "num_distractors": 0,
  "reasoning_depth": 1,
  "input_signature_hash": "..."
}

Loading

from datasets import load_dataset

ds = load_dataset("YOUR_NAMESPACE/arca-synthetic-adaptive")
print(ds)
print(ds["train"][0])

For local loading from the repository root:

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "train": "data/synthetic_adaptive/train.jsonl",
        "validation": "data/synthetic_adaptive/val.jsonl",
        "test_iid": "data/synthetic_adaptive/test_iid.jsonl",
        "test_ood": "data/synthetic_adaptive/test_ood.jsonl",
    },
)

Generation and Validation

Generator:

python3 experiments/synthetic_adaptive/generate_dataset.py

Validator:

python3 experiments/synthetic_adaptive/validate_dataset.py --write-report

Validation checks:

• gold path exists in the original graph
• query traversal reaches target_node
• target_value == node_values[target_node]
• perfect constraints preserve the full gold path
• critical-delete variants delete at least one gold edge
• non-critical-delete preserves gold edges
• false-edge variants add edges not in the original graph
• no input-signature leakage across splits
• target classes are approximately balanced

Dataset Statistics

See dataset_stats.json and validation_report.json.

The default generated dataset has:

• 12 relation labels
• 10 target value classes
• approximately balanced target values
• no input-signature leakage across train/validation/test splits
• IID splits with 2-4 hops and 8-12 nodes
• OOD split with 5-6 hops and 15-20 nodes

Limitations

• This is a symbolic graph benchmark, not a natural-language benchmark.
• It tests path retrieval and constraint robustness, not broad reasoning.
• It is intended for mechanism-level diagnostics, not state-of-the-art claims.
• A model-training benchmark is intentionally not included in this dataset package.