If you thought the AI arms race was all about bigger models and more parameters, think again. The latest battleground for advancing artificial reasoning isn’t just in neural net architecture—it’s in the data itself. Hugging Face, Bespoke Labs, and Together.ai have teamed up to kick off the Reasoning Datasets Competition, a call to arms for researchers, developers, and hobbyists alike to craft innovative datasets that can teach machines to reason better.

Since the release of DeepSeek-R1 in January 2025, reasoning datasets have exploded on Hugging Face’s Hub. Heavy hitters like OpenThoughts-114k, OpenCodeReasoning, and codeforces-cot have primarily focused on domains with clear-cut answers—math, coding, science—where correctness is easy to verify. OpenThoughts-114k alone has been instrumental in training over 230 models.

But the organizers argue that the next wave of breakthroughs won’t come from tweaking model architectures or scaling parameters, but from better data—datasets that reflect the messy, uncertain, and richly nuanced real world. To this end, the competition explicitly encourages datasets that explore novel domains and complex reasoning tasks.

Think legal judgments, financial analysis, literary interpretation, and philosophical ethics. Imagine datasets that push models to handle diagrammatic reasoning, counterfactual thinking, or constraint satisfaction problems like advanced scheduling or resource allocation. They’re even interested in datasets that foster reasoning distillation—transferring reasoning skills from large to smaller models, inspired by insights from the DeepSeek paper.

The competition launches April 9, 2025, with submissions due by May 1 (11:59 PM PT), and winners announced shortly after on May 5. To enter, participants need to upload a proof-of-concept dataset to the Hugging Face Hub, tagged with reasoning-datasets-competition.

Minimum requirements are fairly straightforward:

  • At least 100 examples.
  • A detailed dataset card covering purpose, creation method, example uses, limitations or biases.
  • A viewer preview compatible with Hugging Face’s interface.
  • Clear licensing for research use.

But the organizers are clear: this is your pitch. Go beyond minimums. Make your dataset stand out by showing novelty in approach, domain coverage, and quality of examples. Judges will be looking for datasets that are reasoning-rich, clear, diverse, and minimize hallucination.

To sweeten the deal, the competition offers some tidy incentives:

  • First place bags $1,500 in Together.ai API credits, $1,500 Amazon (or country-specific) gift card, plus Hugging Face Pro subscription and compute credits.
  • Runners-up get $500 gift cards with similar Hugging Face perks.
  • A special spotlight on innovative uses of the Curator tool awards four $250 gift cards.
  • And everyone who participates earns $50 in Together.ai API credits.

The competition isn’t just about direct reasoning datasets. It embraces the broader ecosystem—think datasets that train models to classify reasoning types, detect logical flaws, or evaluate evidence credibility. These meta-level contributions might not need massive resources but could have outsized impact in helping machines sharpen their thinking skills.

And don’t sweat the process too much. While Hugging Face and Bespoke Labs provide example templates, code snippets, and Colab notebooks (including GPU-based and API-only distillation demos), teams can use any tools or methods they like. Synthetic data? Real-world data? LLM-generated? All are welcome.

This competition arrives at an exciting moment. With over 230 models already trained on datasets like OpenThoughts-114k, the community is primed for a new wave of data-driven innovation. The organizers recognize that AI’s reasoning capabilities hinge not just on how models are built, but on the richness and complexity of the datasets that shape them.

If you’ve ever dreamed of creating a dataset that pushes AI’s reasoning boundaries into uncharted territory—be it legal, financial, ethical, or downright quirky—this is your chance to shine, get recognized, and snag some serious resources along the way.

For more details and to register, head over to the Hugging Face discussion thread and join the Discord channel #reasoning-dataset-competition. The future of reasoning isn’t just being coded—it’s being curated. Will your dataset be the next big leap?

The link has been copied!