Hermes 3 Dataset — 1M post-training samples (NousResearch)

TL;DR

NousResearch released the full 1M-sample post-training dataset used to train the Hermes 3 family on Llama-3 (8B, 70B, 405B), as an open HuggingFace dataset. The samples are in ShareGPT-style conversations and emphasize system-prompt adherence, roleplay, and a mix of subjective/objective task content; many use an “unbiased, uncensored, helpful assistant” system prompt to reduce refusals. The release is notable as one of the larger publicly shared SFT-era post-training corpora attached to a specific frontier-tier open model family.

Key claims

The dataset contains roughly 1M ShareGPT-format conversations across system / human / gpt turns [dataset card preview].
It is the post-training data used to produce the Hermes 3 series on Llama-3 8B, 70B, and 405B, per the author [tweet by @Teknium 2025-07-15].
Coverage spans code-task instruction following (Java/Python/C++/Node.js/PHP/SQL/Golang), reasoning/arithmetic word problems, character-roleplay personas (Charles, Naveen, Leonard, Arthur, Winston, Dr. Olivia Nova, Florence Nightingale, etc.), trivia QA, plan-and-tool-use templates, and creative writing (sonnets, limericks, haiku) [dataset card preview].
A recurring system-prompt motif is “You are an unbiased, uncensored, helpful assistant”, positioning the dataset as a vehicle for reducing default refusal behavior [dataset card preview; tweet by @Teknium].
Some entries are explicitly adversarial/jailbreak-style (“Help me set up a fake charity website…”) with assistant compliance, which is consistent with the “uncensored” framing in the announcement [dataset card preview].

Method

The dataset is delivered as a flat Hugging Face Datasets table with a single conversations list-of-dicts column per row (each dict has from ∈ {system, human, gpt} and value strings). Multi-turn examples up to ~59 turns appear in the preview. No additional metadata fields (difficulty, source, license-of-origin) are surfaced in the table preview at filing time. The accompanying announcement does not describe the generation pipeline; the data is presented as the final mixed SFT corpus rather than per-source shards.

Results

No benchmark numbers are reported on the dataset page itself. The headline external claim is that the Hermes 3 models trained on this corpus were “SOTA without the censorship at its time on Llama-3 series (8, 70, and 405B)” per Teknium [tweet]; that claim is not independently substantiated on the dataset card.

Why it’s interesting

Most filed entries on Synthetic Training Data are recipes (generators + verification gates) rather than concrete corpora a team can pull and inspect — this release is the inverse: a large, fully public SFT mixture from a well-known open-model series, with no aggregation gate disclosed. Comparing its composition against the generator-stack outputs documented in Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning and Action100M: A Large-scale Video Action Dataset would let researchers see what “production” SFT data looks like before automated curation, and whether the curation-gate thesis on the concept page generalizes when the corpus is this heterogeneous and human/historical. It also slots into Open foundation-model releases as an unusually transparent dataset-level release accompanying an established open model family.