teknium

--- base_model: mistralai/Mistral-7B-v0.1 tags: - mistral - instruct - finetune - chatml - gpt4 - synthetic data - distillation model-index: - name: OpenHermes-2-Mistral-7B results: [] license: apache-2.0 language: - en datasets: - teknium/OpenHermes-2.5 ---

In the tapestry of Greek mythology, Hermes reigns as the eloquent Messenger of the Gods, a deity who deftly bridges the realms through the art of communication. It is in homage to this divine mediator that I name this advanced LLM "Hermes," a system crafted to navigate the complex intricacies of human discourse with celestial finesse. OpenHermes 2 Mistral 7B is a state of the art Mistral Fine-tune. OpenHermes was trained on 900,000 entries of primarily GPT-4 generated data, from open datasets across the AI landscape. [More details soon] Filtering was extensive of these public datasets, as well as conversion of all formats to ShareGPT, which was then further transformed by axolotl to use ChatML. Huge thank you to WingLian, One, and a16z for compute access and for sponsoring my work, and all the dataset creators and other people who's work has contributed to this project! Follow all my updates in ML and AI on Twitter: https://twitter.com/Teknium1 Support me on Github Sponsors: https://github.com/sponsors/teknium1 Table of Contents 1. Example Outputs - Chat about programming with a superintelligence - Get a gourmet meal recipe - Talk about the nature of Hermes' consciousness - Chat with Edward Elric from Fullmetal Alchemist 2. Benchmark Results - GPT4All - AGIEval - BigBench - Averages Compared 3. Prompt Format 4. Quantized Models Hermes 2 on Mistral-7B outperforms all Nous & Hermes models of the past, save Hermes 70B, and surpasses most of the current Mistral finetunes across the board. Average Score Comparison between Nous-Hermes Llama-2 and OpenHermes Llama-2 against OpenHermes-2 on Mistral-7B: OpenHermes 2 now uses ChatML as the prompt format, opening up a much more structured system for engaging the LLM in multi-turn chat dialogue. System prompts are now a thing that matters! Hermes 2 was trained to be able to utilize system prompts from the prompt to more strongly engage in instructions that span over many turns. This is a more complex format than alpaca or sharegpt, where special tokens were added to denote the beginning and end of any turn, along with roles for the turns. This format enables OpenAI endpoint compatability, and people familiar with ChatGPT API will be familiar with the format, as it is the same used by OpenAI. This prompt is available as a chat template, which means you can format messages using the `tokenizer.applychattemplate()` method: When tokenizing messages for generation, set `addgenerationprompt=True` when calling `applychattemplate()`. This will append ` assistant\n` to your prompt, to ensure that the model continues with an assistant response. To utilize the prompt format without a system prompt, simply leave the line out. Currently, I recommend using LM Studio for chatting with Hermes 2. It is a GUI application that utilizes GGUF models with a llama.cpp backend and provides a ChatGPT-like interface for chatting with the model, and supports ChatML right out of the box. In LM-Studio, simply select the ChatML Prefix on the settings side pane: The Bloke has quantized Open Hermes 2 in GPTQ, GGUF, and AWQ! Available here: https://huggingface.co/TheBloke/OpenHermes-2-Mistral-7B-GPTQ https://huggingface.co/TheBloke/OpenHermes-2-Mistral-7B-GGUF https://huggingface.co/TheBloke/OpenHermes-2-Mistral-7B-AWQ

Collective Cognition v1.1 is a state-of-the-art model fine-tuned using the Mistral approach. This model is particularly notable for its performance, outperforming many 70B models on the TruthfulQA benchmark. This benchmark assesses models for common misconceptions, potentially indicating hallucination rates. Special Features: - Quick Training: This model was trained in just 3 minutes on a single 4090 with a qlora, and competes with 70B scale Llama-2 Models at TruthfulQA. - Limited Data: Despite its exceptional performance, it was trained on only ONE HUNDRED data points, all of which were gathered from a platform reminiscent of ShareGPT. - Extreme TruthfulQA Benchmark: This model is competing strongly with top 70B models on the TruthfulQA benchmark despite the small dataset and qlora training! Special thanks to @a16z and all contributors to the Collective Cognition dataset for making the development of this model possible. The model was trained using data from the Collective Cognition website. The efficacy of this dataset is demonstrated by the model's stellar performance, suggesting that further expansion of this dataset could yield even more promising results. The data is reminiscent of that collected from platforms like ShareGPT. You can contribute to the growth of the dataset by sharing your own ChatGPT chats here. You can download the datasets created by Collective Cognition here: https://huggingface.co/CollectiveCognition - TruthfulQA: Collective Cognition v1.1 has notably outperformed various 70B models on the TruthfulQA benchmark, highlighting its ability to understand and rectify common misconceptions. Training run on wandb here: https://wandb.ai/teknium1/collectivecognition-mistral-7b/runs/collectivecognition-mistral-8/workspace

OpenHermes 13B is the first fine tune of the Hermes dataset that has a fully open source dataset! OpenHermes was trained on 242,000 entries of primarily GPT-4 generated data, from open datasets across the AI landscape, including: - GPTeacher - General Instruct, Roleplay v1, Roleplay v2, and Code Instruct Datasets, by Teknium - WizardLM (v1, evolinstruct 70k), by WizardLM Team/nlpxucan - Airoboros GPT-4 (v1.0), by JonDurbin - Camel-AI's domain expert datasets, by the Camel-AI Team - CodeAlpaca, by Sahil2801 - GPT4-LLM and Unnatural Instructions, by Microsoft Filtering included removal of OpenAI refusals, disclaimers, and "As an AI" type examples and more The base dataset mix the model was trained on is identical to Nous-Hermes', minus the Nous-Instruct and PDACTL datasets which were private datasets. The WANDB Project is public and can be examined at this link: https://wandb.ai/teknium1/openhermes/runs/openhermes-v2-fullft-13b Huge thank you to mainhorse for compute access and a16z for sponsoring my work, and all the dataset creators and other people who's work has contributed to this project! This is a slight improvement on GPT4ALL Suite and BigBench Suite, with a degredation in AGIEval compared to the original hermes. Average Score Comparison between Nous-Hermes Llama-2 and OpenHermes Llama-2: The following hyperparameters were used during training: - learningrate: 2e-05 - trainbatchsize: 2 - seed: 42 - distributedtype: multi-GPU - numdevices: 8 - gradientaccumulationsteps: 8 - totaltrainbatchsize: 128 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: cosine - lrschedulerwarmupsteps: 300 - numepochs: 3

OpenHermes 7B is the first fine tune of the Hermes dataset that has a fully open source dataset! What is unique about this 7B model is that it used sample packing, which speeds up training by many multiples if the dataset token averages arent near the max sequence length. OpenHermes was trained on 242,000 entries of primarily GPT-4 generated data, from open datasets across the AI landscape, including: - GPTeacher - General Instruct, Roleplay v1, Roleplay v2, and Code Instruct Datasets, by Teknium - WizardLM (v1, evolinstruct 70k), by WizardLM Team/nlpxucan - Airoboros GPT-4 (v1.0), by JonDurbin - Camel-AI's domain expert datasets, by the Camel-AI Team - CodeAlpaca, by Sahil2801 - GPT4-LLM and Unnatural Instructions, by Microsoft Filtering included removal of OpenAI refusals, disclaimers, and "As an AI" type examples and more The base dataset mix the model was trained on is identical to Nous-Hermes', minus the Nous-Instruct and PDACTL datasets which were private datasets. The WANDB Project is public and can be examined at this link: https://wandb.ai/teknium1/openhermes/runs/openhermes-v2-qlora-7b-packed Huge thank you to mainhorse for compute access and a16z for sponsoring my work, and all the dataset creators and other people who's work has contributed to this project!