ornstein-curated-100k

# Ornstein Curated 100K **A curriculum-sorted reasoning dataset for SFT and post-training experiments.** Ornstein Curated 100K is a multi-domain instruction dataset built around explicit reasoning traces, difficulty progression, and curriculum-style ordering. The dataset contains 100,000 samples across mathematics, programming, conversational reasoning, and cognitive-science-inspired tasks. It is sorted from easier to harder examples so users can train with the provided order, compare against shuffled baselines, or build their own curriculum schedules. > Ornstein is for users who want more than a pile of prompts. It is a structured reasoning corpus with difficulty metadata, signal scores, and a clear curriculum hypothesis to test. ## Why Use This Dataset? Ornstein is designed for: - supervised fine-tuning of instruct and chat models - curriculum-learning experiments - reasoning trace training - math and code reasoning specialization - comparing ordered vs shuffled SFT - filtering by difficulty, signal score, source, or domain It is especially useful if you want a dataset that is large enough to train on, small enough to inspect, and structured enough to support actual experiments. ## Dataset at a Glance | Property | Value | |---|---:| | Total samples | 100,000 | | Domains | Math, programming, conversational, cognitive science | | Format | JSON Lines with conversation structure | | License | CC-BY 4.0 | | Ordering | Difficulty-sorted, easy to hard | | Mean signal score | 67.3 ± 21.1 | | Mean difficulty | 0.481 | ## Domain Mix  | Domain | Samples | Share | |---|---:|---:| | Mathematics | 40,000 | 40% | | Programming | 30,000 | 30% | | Conversational | 20,000 | 20% | | Cognitive Science | 10,000 | 10% | Mathematics and programming make up the majority of the dataset, with conversational and cognitive-science examples included to broaden the reasoning distribution. ## Quality and Curriculum Signals Ornstein includes metadata intended to support filtering and curriculum experiments. | Metric | Value | |---|---:| | Mean signal score | 67.3 ± 21.1 | | Mean solution length | 287 words | | Self-correction markers | 38% of samples | | Verification markers | 62% of samples | | Curriculum smoothness | 98% variance reduction vs random ordering | The signal score is an automated curation signal. It should be treated as a useful filtering and analysis feature, not as a human preference label. ### Correlation Structure  Signal score, difficulty, and response length show the expected positive relationships for a reasoning-focused curriculum corpus. See `statistical_summary.txt` for the full statistical report. ### Difficulty Progression  The provided order is difficulty-sorted. This makes the dataset useful for curriculum training and for direct ordered-vs-shuffled ablations. ## Schema Each row follows a conversational SFT structure with additional metadata. | Field | Description | |---|---| | `problem` | Original task or prompt | | `solution` | Target assistant response | | `signal_score` | Automated reasoning-signal score | | `difficulty` | Normalized difficulty estimate | | `metadata` | Additional curation metadata | | `source` | Upstream source | | `type` | Domain label | | `conversations` | Chat-format messages | | `code` | Optional code field for programming examples, where present | Example structure: ```json { "conversations": [ {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."} ], "difficulty": 0.65, "signal_score": 72.5, "source": "openmathinstruct2", "type": "math" } ``` ## Usage ```python from datasets import load_dataset ds = load_dataset("DJLougen/ornstein-curated-100k", split="train") print(ds[0]["conversations"]) ``` Use the provided order for curriculum-style training: ```python curriculum_ds = ds ``` Create a shuffled baseline: ```python shuffled_ds = ds.shuffle(seed=42) ``` Filter by domain: ```python math_ds = ds.filter(lambda x: x["type"] == "math") code_ds = ds.filter(lambda x: x["type"] == "code") ``` Filter by difficulty: ```python easy = ds.filter(lambda x: x["difficulty"] < 0.33) medium = ds.filter(lambda x: 0.33 <= x["difficulty"] < 0.66) hard = ds.filter(lambda x: x["difficulty"] >= 0.66) ``` Filter by signal score: ```python high_signal = ds.filter(lambda x: x["signal_score"] >= 70) ``` ## Training Recommendations For curriculum experiments: 1. Train once using the provided dataset order. 2. Train a matched run using a shuffled version. 3. Keep seed, batch size, learning rate, token budget, and chat template constant. 4. Compare downstream reasoning, code, and instruction-following evals. 5. Report both the curriculum result and the shuffled baseline. For standard SFT: - Use the full dataset for broad reasoning coverage. - Use domain filters for math, code, or conversational subsets. - Use difficulty and signal-score filters for curriculum or quality-weighted recipes. - Inspect examples before using the dataset in production training. ## Statistical Documentation The repository includes `statistical_summary.txt` with additional descriptive statistics, correlation tests, ANOVA results, power analysis, and curriculum validation details. ## Attribution Derived from: - [OpenMathInstruct-2](https://huggingface.co/datasets/nvidia/OpenMathInstruct-2), NVIDIA - [OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning), NVIDIA - [Magpie-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered), Magpie-Align - [SCoRe](https://huggingface.co/datasets/jon7009/SCoRe), Structured Chain of Reasoning ## Limitations - Difficulty and signal scores are automated estimates. - The dataset is heavily weighted toward math and programming. - Some examples may contain long reasoning traces that need formatting for your target model. - Curriculum ordering is a training hypothesis, not a guaranteed improvement. - Users should run downstream evaluations after training. ## Citation ```bibtex @dataset{ornstein_curated_2025, author = {Lougen, Daniel}, title = {Ornstein Curated 100K: A Curriculum-Sorted Multi-Domain Reasoning Dataset}, year = {2025}, publisher = {Hugging Face}, url = {https://huggingface.co/datasets/DJLougen/ornstein-curated-100k} } ```