Back to Gap Map Model components / Training & synthetic datasets

Hermes 3 Dataset

open / Overall score: 2.5

Nous Research's release of the instruction and chat data behind Hermes 3, a mixture of curated and generated examples. A significant portion of it was retained in the Hermes 4 training data, and the Hermes 3 technical report describes how the mixture was built.

Openness

5 high confidence

5.0

license: Apache-2.0
gated: false
dataset_card: present

Apache-2.0, ungated, dataset card.

https://huggingface.co/api/datasets/NousResearch/Hermes-3-Dataset recorded 2026-10-08
Hub metadata for "NousResearch/Hermes-3-Dataset" reports the dataset as not gated and carries the tag "license:apache-2.0".
https://huggingface.co/datasets/NousResearch/Hermes-3-Dataset recorded 2026-10-08
The dataset page, titled "NousResearch/Hermes-3-Dataset · Datasets at Hugging Face", carries the tag "license:apache-2.0", reports the repository as not gated and has a dataset card.

Adoption

1 high confidence

1.0

Adoption is measured by Hugging Face downloads of the dataset over the trailing month, the only usage channel on record for the Hermes 3 Dataset.

https://huggingface.co/api/datasets/NousResearch/Hermes-3-Dataset recorded 2026-08-13
688 downloads in the trailing 30 days for NousResearch/Hermes-3-Dataset

Capability

4 high confidence

4.0

Proven value with documented Hermes 3 results and explicitly folded into Hermes 4 (2025), but independent reuse is limited to community quantizers. Table 5 of the Hermes 3 report puts Hermes 3 ahead of Llama 3.1 Instruct on the majority of its seventeen benchmarks at all three sizes.

https://arxiv.org/pdf/2408.11857 recorded 2026-08-13
Hermes 3 beats Llama 3.1 Instruct on most benchmarks at 405B/70B/8B
https://arxiv.org/pdf/2508.18255 recorded 2026-08-13
Hermes 4 report: a significant portion of the Hermes 3 dataset was retained

Verified 2026-09-27, except some scores, still awaiting re-confirmation