QVAC Genesis III Releases 191.43B Tokens of STEM Reasoning Data for Small AI Models
Tether AI Research released QVAC Genesis III on September 23, 2026, as an open synthetic training corpus containing 191.43 billion tokens across 159.6 million documents and 19 STEM domains. The dataset is available through Hugging Face under CC BY-NC 4.0, which permits sharing and adaptation with attribution for non-commercial purposes.
Genesis III targets smaller language models intended for resource-constrained and local deployments. Its associated research evaluates models with 1.7 billion parameters trained from scratch, allowing the data-generation strategy to be compared under controlled model and token budgets instead of attributing results to a larger model.
The published experiments report gains over Cosmopedia-v2-trained baselines on ARC, GPQA Diamond and MMLU STEM. The strongest reported differences include +28.57 percentage points on ARC-Easy and +21.35 points on ARC-Challenge. An Option-Level-trained 1.7B model reached a 99.45% valid-answer rate in the reported evaluation. These figures come from the Genesis III research team's controlled experiments and should be read as dataset-ablation results, not general-purpose rankings of production models.
What is in Genesis III
Genesis III spans 19 curriculum-aligned STEM areas across high-school, college and professional difficulty levels. Tether lists subjects including biology, chemistry, physics, mathematics, computer science, medicine, astronomy, electrical engineering, statistics and machine learning.
The release expands the QVAC Genesis series substantially. Genesis I introduced a smaller STEM corpus in October 2025, followed by Genesis II in December 2025. Genesis III is the largest published release in the series at 191.43 billion tokens.
The dataset can be streamed through the Hugging Face datasets library without preparing the full corpus locally:
from datasets import load_dataset
ds = load_dataset("qvac/GenesisIII", split="train", streaming=True)
print(next(iter(ds)))
Researchers can also download the repository locally with the Hugging Face CLI for preprocessing and training workflows that require a fixed local copy.
Two synthetic-data strategies
The Genesis III paper describes two complementary generation strategies designed around the behavior of a smaller student model.
Failure Analysis starts from questions the student answers incorrectly. A stronger teacher model produces a corrected explanation and answer, concentrating training material around weaknesses exposed by the smaller model.
Option-Level reasoning expands successful multiple-choice examples into explanations covering the correct answer and the alternatives. This creates explicit reasoning about why each option is valid or invalid instead of retaining only the final label.
This student-conditioned construction is the central research idea behind the corpus. Training data is selected and expanded according to where a small model needs additional reasoning signal, rather than generated as a uniform collection of unrelated synthetic STEM text.
What the 1.7B experiments show
The paper uses controlled from-scratch 1.7B-parameter training runs to compare Genesis III with open synthetic-data baselines. Evaluations include ARC-Easy, ARC-Challenge, GPQA Diamond and MMLU STEM subsets.
The authors report that Genesis III-trained models consistently improve over their Cosmopedia-v2 controls across the evaluated STEM reasoning tasks. The reported maximum improvements of 28.57 points on ARC-Easy and 21.35 points on ARC-Challenge are useful evidence that targeted synthetic reasoning data can improve learning efficiency under a constrained model budget.
The experiment establishes a narrower result than a general model ranking: under the paper's controlled training setup, changing the training corpus materially changes STEM reasoning performance without increasing the tested model size.
The research paper was published on September 17 and Tether says it has been accepted for presentation at COLM 2026.
License and deployment boundary
Genesis III is published under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. The license permits sharing and adaptation with attribution when the use is non-commercial.
Teams evaluating the corpus should assess the purpose of the intended use against the license's NonCommercial definition. Non-commercial research, education and experimentation can fit the published grant; uses primarily intended for commercial advantage or monetary compensation require separate rights or permission covering the intended use.
The dataset is an input to model training rather than a ready-to-run local model. Training even a 1.7B model from scratch over a corpus of this scale remains a substantial compute workload. The local-AI relevance comes from the research objective: improving the capability learned by smaller models that can later be deployed on lower-resource hardware.
Where Genesis III is most useful
The corpus is most relevant to researchers building compact STEM models, studying synthetic-data efficiency, testing curriculum-oriented pretraining or reproducing data-ablation experiments. The public dataset and paper make the training-data side of the experiment inspectable, while the fixed 1.7B test scale provides a practical reference point for follow-on work.
For local-AI builders, the broader implication is data efficiency. Smaller deployable models depend heavily on the quality and targeting of their training tokens. Genesis III provides a large public corpus for testing whether student-conditioned synthetic reasoning can preserve more STEM capability within a constrained parameter budget.
Sources
- Tether AI Research — Genesis III release announcement: https://tether.io/news/tether-releases-genesis-iii-to-train-ai-to-explain-not-just-answer-stem-problems/
- QVAC — Genesis dataset page and Hugging Face loading instructions: https://qvac.tether.io/products/genesis
- QVAC Genesis III paper — technical methodology and controlled 1.7B experiments: https://huggingface.co/papers/2609.19513
- Hugging Face — QVAC Genesis III dataset: https://huggingface.co/datasets/qvac/GenesisIII
- Creative Commons — CC BY-NC 4.0 license terms: https://creativecommons.org/licenses/by-nc/4.0/