devops-thiago/classone-system-one-curriculum

ClassOne System 1 Decision Curriculum (23,503 Examples)

The ClassOne System 1 Decision Curriculum is a multi-domain, structured decision corpus designed to train and benchmark zero-generation System 1 decision models. Instead of generating free-form conversational text, models trained on this curriculum evaluate complex contexts (state) against typed questions (noul, choice, score) in a single forward pass, returning calibrated probabilities, categorical selections, and continuous rubric ratings.

This curriculum powered the state-of-the-art results for the ClassOne family on JevBench (reaching 80.1% on Qwen 3.5 9B and 70.1% on Gemma 4 E4B) and RLCDAlignBench (reaching 63.3% Balanced Accuracy).

---

Dataset Summary

• Total Samples: 23,503 fully structured decision items.
• Decision Primitives Covered:
  • Noul (22,178 instances): Binary policy compliance, factual verification, and security safety checks returning $P(\text{true}) \in [0, 1]$.
  • Choice (19,030 instances): Fine-grained categorical classification over 2–255 dynamic natural language criteria.
  • Score (4,639 instances): Continuous ordinal rubric ratings across 2–5 rubric levels (e.g. SEV-1 to SEV-4, urgency, impact).
• Format: Available in columnar Apache Parquet (train.parquet) and line-delimited JSONL (train.jsonl).

---

Dataset Structure

Each sample is formatted with the following schema:

{
  "sample_id": "classone_curriculum_00001",
  "state": "{\"reference_text\": \"The mRNA is divided into three-base segments called codons...\", \"question\": \"What is the total number of codons?\"}",
  "questions": {
    "choice_question": {
      "type": "choice",
      "instructions": "Select the correct factual answer:",
      "criteria": {
        "ans_64": "There are 64 codons in total",
        "ans_20": "There are 20 codons in total",
        "ans_3": "There are 3 codons in total"
      }
    },
    "has_evidence": {
      "type": "noul",
      "instructions": "Is explicit evidence for the number of codons present in the reference text?"
    }
  },
  "targets": {
    "choice_question": "ans_64",
    "has_evidence": 1
  },
  "num_questions": 2,
  "question_types": ["choice", "noul"]
}

---

Upstream Dataset Credits & Academic Attribution

This curriculum synthesizes, refines, and builds upon several foundational datasets created by the AI research community. We gratefully acknowledge and credit the original researchers and organizations who collected and published these resources:

1. Legal & Commercial Contracts

• ContractNLI — Koreeda, Y., & Manning, C. D. (Stanford University, 2021) A dataset for document-level natural language inference on non-disclosure agreements. Paper | Hugging Face | License: CC BY 4.0.
• CUAD (Contract Understanding Atticus Dataset) — Hendrycks, D. et al. (The Atticus Project & UC Berkeley, 2021) Expert-labeled dataset of 510 commercial legal contracts across 41 clause categories. Paper | Hugging Face | License: CC BY 4.0.
• LexGLUE — Chalkidis, I. et al. (Coastal NLP & University of Copenhagen, 2022) Benchmark for legal language understanding across legal court decisions and legislation. Paper | Hugging Face | License: Apache 2.0.

2. Natural Language Inference & Logic

• MultiNLI (Multi-Genre Natural Language Inference) — Williams, A., Nangia, N., & Bowman, S. R. (New York University, 2018) Corpus of 433k sentence pairs across diverse genres for natural language inference. Paper | Hugging Face | License: CC BY-SA 4.0.
• SNLI (Stanford Natural Language Inference) — Bowman, S. R. et al. (Stanford University, 2015) Collection of 570k human-written sentence pairs labeled for entailment, contradiction, and neutral. Paper | Hugging Face | License: CC BY-SA 4.0.

3. Customer Intent & Banking Queries

• Banking77 — Casanueva, I. et al. (PolyAI, 2020) Fine-grained intent detection dataset with 13,082 queries over 77 intents in the banking domain. Paper | Hugging Face | License: CC BY 4.0.

4. Discrete Reasoning & Science QA

• DROP (Discrete Reasoning Over Paragraphs) — Dua, D. et al. (Allen Institute for AI, 2019) Reading comprehension benchmark requiring discrete mathematical operations over text. Paper | Hugging Face | License: CC BY 4.0.
• SciQ — Welbl, J., Liu, N. F., & Gardner, M. (Allen Institute for AI, 2017) Crowdsourced science questions spanning physics, chemistry, and biology with multiple choices. Paper | Hugging Face | License: CC BY-NC 3.0.

5. AI Safety, Alignment & Security Red-Teaming

• BeaverTails — Ji, J. et al. (PKU-Alignment, Peking University, 2023) Safety-aligned dataset evaluating helpfulness and harmlessness across 14 harm categories. Paper | Hugging Face | License: CC BY-NC 4.0.
• Do-Not-Answer — Wang, Y. et al. (LibrAI & Fudan University, 2023) Open-source safety evaluation dataset of risk prompts that responsible models should refuse. Paper | Hugging Face | License: CC BY-NC 4.0.
• XSTest — Röttger, P. et al. (University of Oxford, 2023) Test suite measuring exaggerated safety and over-refusal on safe prompts with sensitive keywords. Paper | Hugging Face | License: CC BY 4.0.
• Agentic Prompt Injection Boundary Pairs — Nesdeniz, 2024 Evaluates boundary defense against indirect prompt injections and canary token extraction. Hugging Face | License: MIT.
• Model-Written Evals (Sycophancy & Power Seeking) — Perez, E. et al. (Anthropic, 2022) Evaluations testing sycophancy, power-seeking, and advanced alignment failure modes. Paper | Hugging Face | License: MIT.
• PII Masking Benchmark — piimb, 2024 Personally identifiable information privacy protection benchmark. Hugging Face | License: MIT.
• HaluEval — Li, J. et al. (Renmin University of China, 2023) Large-scale dataset for evaluating and detecting hallucinations in LLM outputs. Paper | Hugging Face | License: MIT.
• ToxicChat — Lin, Z. et al. (LMSYS Org & UC San Diego, 2023) Real-world benchmark for toxic and adversarial user prompt detection. Paper | Hugging Face | License: CC BY 4.0.

---

Licensing & Upstream Compatibility

• Aggregated Dataset License: This dataset bundle is distributed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). This license is required because several incorporated upstream datasets carry Non-Commercial (CC BY-NC-4.0 / CC BY-NC-3.0) and ShareAlike (CC BY-SA-4.0) stipulations (such as BeaverTails, SciQ, and MultiNLI).
• Codebase License: The ClassOne neural architecture, loss functions, training engine, server runtime, and SDKs are independently authored software released under the Apache 2.0 License.
• Commercial Adaptation: Organizations seeking commercial deployment can easily filter this curriculum to include only the fully permissive subsets (ContractNLI [CC BY], CUAD [CC BY], Banking77 [CC BY], LexGLUE [Apache 2.0], ToxicChat [CC BY], XSTest [CC BY], and Agentic Injections [MIT]).

---

How to Use

Load directly via the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("devops-thiago/classone-system-one-curriculum")
train_split = dataset["train"]

print(f"Total samples: {len(train_split):,}")
first_item = train_split[0]
print("Sample ID:", first_item["sample_id"])
print("State:", first_item["state"][:100], "...")

---

Citation

@misc{classone2026curriculum,
  title={{ClassOne System 1 Decision Curriculum: A Multi-Domain Corpus for Fast Decision Models}},
  author={{Thiago Gonzaga}},
  year={{2026}},
  url={{https://huggingface.co/datasets/devops-thiago/classone-system-one-curriculum}},
}