Arxh-DeepLogic is designed to test whether an AI system can combine deep-learning knowledge with multi-step reasoning rather than relying only on definition recall.

What is inside?

1,000 unique synthetic examples

Split	Examples
train	800
validation	100
test	100

The benchmark spans architecture parameters and tensor geometry, attention, optimization, gradient flow, normalization and regularization, generalization, data quality, debugging, memory and serving systems, causal inference, research methodology, counterfactual reasoning, and hard-constraint selection.

Reasoning curriculum

Examples are tagged with curriculum levels:

L1 → foundational inference
L2 → direct multi-step reasoning
L3 → technical diagnosis
L4 → causal/architectural reasoning
L5 → counterfactual and systems reasoning
L6 → research-grade constraint/hypothesis reasoning

Why this dataset exists

Many technical datasets can test whether a model knows a concept. Arxh-DeepLogic is aimed at the harder question:

Can the model use several technical facts at once, respect constraints, distinguish evidence from causation, and reach a verifiable conclusion?

Data format

Each record contains:

question_id
domain
task_type
difficulty
curriculum_level
question
options
answer_type
answer
answer_index
concise_explanation
required_concepts
constraints
verifier
reproducibility metadata

The full schema is in configs/schema.json.

Loading

from datasets import load_dataset

ds = load_dataset("YOUR_HF_NAMESPACE/Arxh-DeepLogic")
print(ds["train"][0])

Hugging Face infers the train, validation, and test splits from the filenames.

Blind evaluation

For a benchmark-style evaluation, use:

benchmark/test_public.jsonl — questions only
benchmark/test_reference.jsonl — gold labels

The reference file is intended to stay separate from the submission workflow.

Example prediction format:

{"question_id":"ADL-0.1.2-0912","prediction":"4x"}

Evaluate with:

python scripts/evaluate_predictions.py --predictions predictions.jsonl

Validation

python scripts/validate_dataset.py

Data generation and quality

v0.1.2 is synthetic and programmatically generated with a deterministic seed and internally validated structural checks. It is not presented as human-expert audited.

The project is intentionally open about this limitation. A future research-grade release should add independent expert review, adversarial question audits, contamination testing, and external benchmark validation.

Intended use

evaluating reasoning-oriented language models
curriculum learning experiments
supervised fine-tuning experiments
technical reasoning probes
research prototypes for AI assistants and agents

Limitations

A high score on Arxh-DeepLogic does not establish broad intelligence, real-world reliability, or scientific competence. Synthetic generation can introduce template artifacts, distributional regularities, and imperfect assumptions.

Citation

@misc{arxh_deeplogic_2026,
  title        = {Arxh-DeepLogic v0.1.2: Deep Learning and Advanced Logical Reasoning Benchmark},
  author       = {{Arxh AI}},
  year         = {2026},
  version      = {0.1.2},
  publisher    = {Arxh AI},
  license      = {CC BY 4.0}
}

Release

Arxh-DeepLogic v0.1.2
Publisher: Arxh AI

Downloads last month: 57

Total file size:

829 kB