openhermes

**OpenHermes** is a **highly influential family of fine-tuned language models created by Teknium that consistently tops open-source leaderboards for 7B-class models** — trained on the OpenHermes-2.5 dataset (1 million+ high-quality conversations aggregated from OpenOrca reasoning traces, Airoboros creative writing, CamelAI domain knowledge, and GPT-4 synthetic data), producing uncensored, instruction-following models that serve as the base for many community model merges and fine-tunes. **What Is OpenHermes?** - **Definition**: A series of fine-tuned language models (primarily based on Mistral-7B) created by Teknium — an independent AI researcher known for producing some of the highest-quality open-source fine-tunes through careful dataset curation and training methodology. - **OpenHermes-2.5 Dataset**: The key innovation is the training dataset — a massive aggregation of 1M+ conversations from multiple high-quality sources: OpenOrca (reasoning traces from GPT-4), Airoboros (creative writing and roleplay), CamelAI (domain-specific knowledge), and GPT-4 synthesis (high-quality synthetic conversations). - **Uncensored Philosophy**: OpenHermes models are trained without heavy safety filtering — following the philosophy that the model should be capable and the application layer should handle content policy, giving developers full control over model behavior. - **Leaderboard Performance**: OpenHermes models (especially OpenHermes-2.5-Mistral-7B) consistently rank at or near the top of the Hugging Face Open LLM Leaderboard for the 7B parameter class — outperforming many larger models on reasoning benchmarks. **Why OpenHermes Matters** - **Data Quality Over Model Size**: OpenHermes demonstrates that a well-curated training dataset matters more than model size — a 7B model trained on high-quality data outperforms 13B and even some 70B models trained on lower-quality data. - **Community Foundation**: OpenHermes models serve as the base for hundreds of community model merges — the "Hermes" lineage appears in many of the most popular merged models on Hugging Face. - **Reasoning Strength**: The inclusion of OpenOrca reasoning traces (step-by-step problem solving from GPT-4) gives OpenHermes models unusually strong reasoning capabilities for their size. - **Practical Instruction Following**: OpenHermes models excel at following complex, multi-step instructions — making them practical for real-world applications beyond benchmark performance. **OpenHermes is the fine-tuned model family that proved dataset curation is the key to open-source model quality** — by aggregating 1M+ high-quality conversations from diverse sources into the OpenHermes-2.5 dataset, Teknium created 7B models that rival much larger competitors and serve as the foundation for the community's most popular model merges.