YouAI Data — stem-reasoning-v1.0.0-ccbysa-001

Dataset Description

YouAI Data — 1,000 STEM training examples extracted from verified CC-BY-SA expert sources — real domain experts solving real problems, not synthetic LLM generation. Contains 394 step-by-step reasoning chains and 569 instruction/response pairs across 332 unique sources. Every example traces to a source URL, available source metadata, and verified license. Includes 364 DPO preference pairs as a free companion dataset. Includes 469 raw reasoning examples with explicit answer labels suitable for RLVR; the remaining rows are exported in chat/SFT formats only.

Dataset Summary

• Size: 1,000 examples
• Domain: STEM
• Formats: ShareGPT, Alpaca, ChatML, DPO Pairs, Raw Reasoning (RLVR-ready rows with explicit answer labels) — JSONL, with matching Parquet for non-empty shards
• License: CC-BY-SA-4.0
• Created on: 2026-04-11 19:44:15 UTC

Quality Metrics

• Mean quality score: 4.55 / 5.0
• Min quality score: 4.46
• Max quality score: 4.81
• Verification: YouAI multi-stage verification on every example

Diversity Metrics

• Unique sources: 332
• Unique concepts tagged: 2,881
• Difficulty: introductory: 348 (35%) · intermediate: 170 (17%) · advanced: 482 (48%)
• No single source exceeds 15% of the dataset

What's Included

• 5 export variants: ShareGPT · Alpaca · ChatML · DPO Pairs · Raw Reasoning
• DPO preference pairs: 364 pairs; best-scoring = chosen, lowest-scoring = rejected
• Raw reasoning rows: 469 rows with explicit answer labels suitable for RLVR
• Curriculum splits: train-introductory, train-intermediate, train-advanced
• Sidecar provenance: Full metadata per example in sidecar/dataclaw_meta.jsonl

Verification & Quality Assurance

Every example in this dataset passes a multi-stage review process developed by YouAI:

1. Automated filtering — Raw content is scored and filtered for information density, coherence, and domain relevance. Typically 30–40% of raw material is discarded at this stage.
2. Multi-answer validation — Multiple independent answers are generated per question. Examples are only retained when answers reach a minimum agreement threshold; conflicting answers are discarded.
3. Computational verification — Where applicable (mathematics, code), answers are verified programmatically against ground truth.
4. YouAI expert review — A structured human expert review pass is applied to sampled examples across every domain, checking for factual accuracy, logical soundness, and format correctness.
5. 4-signal quality scoring — Each example carries a composite quality score derived from multiple independent signals. Only examples meeting the minimum threshold are included.

The specific methods, tooling, and thresholds used in our verification pipeline are proprietary to YouAI.

Decontamination

Decontaminated against GSM8K, MATH, MMLU, HumanEval, MBPP, ARC-Challenge.

Method: 13-gram overlap detection. Zero benchmark test-set overlap.

Source & Provenance

Every example traces to a specific source with a full audit trail:

• Source URL, title, channel, and license — in provenance.json (per-source summary with timestamp range)
• Per-example source metadata — in sidecar/dataclaw_meta.jsonl under source.*
• Available timestamp range (source.timestamp_start, source.timestamp_end in seconds) — in sidecar/dataclaw_meta.jsonl
• Quality score breakdown per signal (quality.breakdown) — in sidecar/dataclaw_meta.jsonl

All source material is verified as commercially licensable prior to inclusion. No sources with NonCommercial (NC) or NoDerivatives (ND) restrictions are used.

Data Collection

YouAI Data uses a proprietary autonomous pipeline for discovery, extraction, and quality assurance. The specific discovery methods, sources, and tooling are not disclosed. What we can share:

• All content originates from openly licensed or public domain sources
• Every example is traceable to its original source URL and license
• The pipeline is designed for domain-expert content — university lectures, technical documentation, professional proceedings — not synthetic generation

How to Use

from datasets import load_dataset

# Load default (ShareGPT, train split)
ds = load_dataset("YouAIData/stem-reasoning-v1.0.0-ccbysa-001", split="train")

# Load specific format
ds_alpaca = load_dataset("YouAIData/stem-reasoning-v1.0.0-ccbysa-001", data_files="data/formats/alpaca.parquet")

# Load DPO preference pairs (when present)
ds_dpo = load_dataset("YouAIData/stem-reasoning-v1.0.0-ccbysa-001", data_files="data/dpo/train-preferences.parquet")

# Load curriculum split
ds_hard = load_dataset("YouAIData/stem-reasoning-v1.0.0-ccbysa-001", data_files="data/curriculum/train-advanced.parquet")

Limitations & Biases

English only (v1). Content skews toward university-level formal register. Speaker demographics reflect availability of open-license expert sources.

Example

A real row sampled from this dataset (ShareGPT format):

{
  "conversations": [
    {
      "from": "human",
      "value": "Derive the Hamiltonian for a spherical pendulum of mass m and length ℓ in a gravitational field g, expressing it in terms of the conjugate momenta P_θ and P_φ."
    },
    {
      "from": "gpt",
      "value": "**Step 1:** The Hamiltonian H is defined as H = P_θ θ̇ + P_φ φ̇ - L, where L is the Lagrangian.\n\n**Step 2:** The conjugate momenta are defined as P_θ = ∂L/∂θ̇ and P_φ = ∂L/∂φ̇.\n\n**Step 3:** From the Lagrangian, P_θ = mℓ² θ̇ and P_φ = mℓ² sin²θ φ̇.\n\n**Step 4:** Solving for θ̇ and φ̇, we get θ̇ = P_θ/(mℓ²) and φ̇ = P_φ/(mℓ² sin²θ).\n\n**Step 5:** Substituting these into the kinetic energy T = (1/2)mℓ² θ̇² + (1/2)mℓ² sin²θ φ̇², we get T = P_θ²/(2mℓ²) + P_φ²/(2mℓ² sin²θ).\n\n**Step 6:** The potential energy V = -mgℓ cosθ.\n\n**Step 7:** Combining T and V, the Hamiltonian is H = P_θ²/(2mℓ²) + P_φ²/(2mℓ² sin²θ) - mgℓ cosθ.\n\n\n**Answer:** H = \\frac{P_{\\theta}^{2}}{2m\\ell^{2}} + \\frac{P_{\\varphi}^{2}}{2m\\ell^{2}\\sin^{2}\\theta} - mg\\ell \\cos\\theta"
    }
  ]
}

Citation

@misc{youaidata_stem_reasoning_v1.0.0_ccbysa_001,
  title={{stem-reasoning-v1.0.0-ccbysa-001}},
  author={{YouAI Data}},
  year={{2026}},
  url={{https://huggingface.co/datasets/YouAIData/stem-reasoning-v1.0.0-ccbysa-001}}
}

Training Results — Demonstrated Improvement

We fine-tuned Qwen3.5-4B (Instruct) using our STEM 001 10K dataset (~10K examples + ~4K DPO pairs) with LoRA to measure the impact on reasoning benchmarks. The results speak for themselves:

Benchmark	Baseline	+ SFT	+ SFT + DPO	Improvement
GSM8K	70.5%	86.0%	87.5%	+17.0%
MATH-500	28.0%	50.0%	54.0%	+26.0%
ARC-Challenge	88.5%	87.5%	89.5%	+1.0%

Training configuration:

• Method: LoRA (r=32, alpha=64) — base model weights frozen
• SFT: 1 epoch, lr=5e-5, batch=2, grad_accum=8, packing enabled
• DPO: 200 steps, lr=2e-5, beta=0.1, implicit PEFT reference model
• Hardware: Single RTX 4090 (24GB), ~2h SFT + ~2.5h DPO
• Data: ~10K SFT examples (ShareGPT format) + ~4K DPO preference pairs from our STEM Reasoning v1 dataset

Key takeaways:

• MATH-500 nearly doubled (+26%) — our multi-step reasoning chains directly transfer to competition-level math
• GSM8K jumped +17% — grade-school math improved dramatically with real expert reasoning examples
• ARC held steady — no catastrophic forgetting on science reasoning; DPO even recovered the slight SFT dip
• DPO adds value on top of SFT — every benchmark improved or held from SFT to SFT+DPO

These results demonstrate that a single, carefully curated 10K-example dataset from real expert reasoning outperforms what many achieve with 100K+ synthetic examples.

Commercial Licensing & Exclusive Datasets

Interested in purchasing our full 10K+ CC-BY datasets with exclusive licensing rights? We offer:

• Standard datasets — Full verified datasets in 5 formats (ShareGPT, Alpaca, ChatML, DPO, Raw Reasoning rows with explicit answer labels)
• Exclusive licensing — 90-day exclusivity on specific domain datasets
• Custom datasets — Tailored to your domain, format, and volume requirements

Contact us:

• HuggingFace: YouAIData
• X / Twitter: @YouAIDATA