988tg/repo-name

Dataset Card for SynthLogic-Instruct

Dataset Details

Dataset Description

SynthLogic-Instruct is a synthetically generated instruction-following dataset designed to enhance the multi-step logical reasoning and code-generation capabilities of large language models. The dataset consists of high-quality prompt-response pairs spanning algorithmic problem solving, mathematical deduction, and structured data analysis tasks. All samples were generated via a verified pipeline using teacher models with rigorous self-consistency filtering and automated unit-test validation for code-related entries.

• Curated by: Open Research Collective (Anonymous Review Track)
• Funded by [optional]: Community Compute Grant Program
• Shared by [optional]: Anonymous Hugging Face Organization
• Language(s) (NLP): English
• License: CC BY 4.0

Dataset Sources [optional]

• Repository: https://github.com/open-research-collective/synthlogic-instruct
• Paper [optional]: Under double-blind review; preprint available upon acceptance
• Demo [optional]: https://huggingface.co/spaces/anonymous-org/synthlogic-eval

Uses

Direct Use

• Fine-tuning base LLMs for improved chain-of-thought reasoning and code synthesis.
• Serving as an evaluation benchmark for measuring logical consistency in instruction-following models.
• Studying the transferability of synthetic reasoning traces to real-world downstream tasks.
• Data augmentation for low-resource reasoning domains where human-annotated data is scarce.

Out-of-Scope Use

• Not intended for training conversational chatbots requiring empathetic or open-ended dialogue.
• Should not be used as ground truth for safety-critical systems (e.g., medical diagnosis, legal advice).
• Code solutions are pedagogical examples and must not be deployed directly in production without independent verification.
• Not suitable for evaluating factual knowledge recall, as all content is synthetically derived.

Dataset Structure

The dataset is stored in Parquet format with the following schema:

Field	Type	Description
id	string	Unique sample identifier
category	string	Task type: algorithm, math, data_analysis, debugging
instruction	string	Natural language prompt describing the task
response	string	Model-generated solution with embedded chain-of-thought
code_snippet	string | null	Executable Python code (present only for code-generation categories)
test_cases	list[dict]	Input-output pairs for automated validation (null for non-code categories)
complexity_level	string	easy, medium, hard based on step count and dependency depth
verification_pass	bool	Whether the response passed all automated consistency checks

Splits: train (45,000 samples), validation (5,000 samples). No test split is released to prevent benchmark contamination.

Dataset Creation

Curation Rationale

Existing instruction datasets often lack sufficient density of verifiable multi-step reasoning traces. Human annotation of such data is expensive and inconsistent. SynthLogic-Instruct addresses this gap through a fully reproducible synthetic pipeline that prioritizes correctness over volume, ensuring every included sample has been mechanically validated.

Source Data

Data Collection and Processing

1. Seed Prompts: 2,000 seed instructions were sampled from public domain problem repositories (Project Euler, Rosetta Code, MIT OCW) under permissive licenses.
2. Generation Pipeline: A teacher model generated 5 candidate responses per seed prompt at temperature=0.8.
3. Filtering: Only samples where ≥4/5 candidates produced semantically equivalent final answers were retained (self-consistency threshold).
4. Code Validation: All code snippets were executed in sandboxed Docker containers against provided test cases; failures were discarded.
5. Deduplication: MinHash-based near-duplicate removal (threshold=0.85) reduced redundancy by 12%.
6. Tools Used: Python 3.11, Ray Data, vLLM, pytest, custom validation framework.

Who are the source data producers?

All data was synthetically generated. Seed prompts originate from publicly licensed educational and programming resources. No individual contributors are identifiable.

Annotations [optional]

Annotation process

No human annotation was performed. Quality assurance was entirely automated:

• Self-consistency voting across multiple generations.
• Unit test execution for code correctness.
• Perplexity-based fluency filtering (samples >2σ above mean perplexity were removed).
• Format compliance checking via regex validators.

Inter-annotator agreement metrics are not applicable due to the synthetic nature of the dataset.

Who are the annotators?

Not applicable. All quality signals are derived from automated verification procedures.

Personal and Sensitive Information

The dataset contains no personal information, PII, or sensitive attributes. All content is synthetically generated from public-domain seeds. No demographic data, identities, or private communications are present.

Bias, Risks, and Limitations

• Synthetic Distribution Shift: Models trained exclusively on this data may exhibit reduced performance on naturally written human queries due to stylistic differences in synthetic reasoning traces.
• Domain Narrowness: Coverage is limited to STEM-oriented reasoning tasks; humanities, creative writing, and social reasoning are absent.
• Teacher Model Bias: Generated content inherits biases and failure modes of the underlying teacher model, including potential over-reliance on specific solution patterns.
• Verification Gaps: Automated tests cannot guarantee semantic correctness for all mathematical proofs; some logically flawed but syntactically valid responses may persist.
• English-Only: No multilingual coverage limits applicability to non-English reasoning evaluation.

Recommendations

Users should mix SynthLogic-Instruct with diverse human-written instruction data during fine-tuning to mitigate distribution shift. Always validate model outputs independently when deploying in applied settings. Researchers are encouraged to audit subset performance by category and complexity_level before drawing broad conclusions about reasoning capability improvements.

Citation [optional]

BibTeX:

@misc{synthlogic2025,
  title={SynthLogic-Instruct: Verified Synthetic Reasoning Traces for Instruction Tuning},
  author={Open Research Collective},
  year={2025},
  note={Under double-blind review}
}

APA:

Open Research Collective. (2025). SynthLogic-Instruct: Verified synthetic reasoning traces for instruction tuning. Under double-blind review.

Glossary [optional]

• Self-Consistency Filtering: A verification method where multiple independent generations are compared; only prompts yielding consistent outputs across samples are retained.
• Chain-of-Thought (CoT): Intermediate reasoning steps explicitly articulated before arriving at a final answer.
• Sandboxed Execution: Isolated runtime environment preventing network access or filesystem modification during code validation.

More Information [optional]

• Full generation pipeline code and configuration: GitHub Repository
• Reproducibility checklist and compute specifications included in repository README
• Community contributions welcome via pull requests for additional categories or validation improvements

Dataset Card Authors [optional]

Members of the Open Research Collective (names withheld for anonymous review).

Dataset Card Contact

For questions or issues, please open a GitHub issue at https://github.com/open-research-collective/synthlogic-instruct/issues or contact synthlogic-support@openresearchcollective.org. ```