--- base_model: teknium/OpenHermes-2.5-Mistral-7B license: apache-2.0 datasets: - teknium/openhermes - argilla/ultrafeedback-binarized-preferences - Intel/orca_dpo_pairs language: - en library_name: transformers pipeline_tag: text-generation --- # DPOpenHermes 7B ![image/png](https://huggingface.co/openaccess-ai-collective/DPOpenHermes-7B/resolve/main/assets/dpopenhermes.png) ## OpenHermes x Notus x Neural This is an RL fine tuned model of [Teknium](https://huggingface.co/teknium)'s [OpenHermes-2.5-Mistral-7B](https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B) using the [Intel/orca_dpo_pairs](https://huggingface.co/datasets/Intel/orca_dpo_pairs) and [argilla/ultrafeedback-binarized-preferences](https://huggingface.co/datasets/argilla/ultrafeedback-binarized-preferences) preference datasets for reinforcement learning using Direct Preference Optimization (DPO) DPOpenHermes is trained using qLoRA. The adapter is also provided in this model repo. # Training Details [Built with Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) DPOpenHermes was trained on a single H100 80GB hosted on RunPod for ~10h for 0.6 epochs of the dataset. https://wandb.ai/oaaic/openhermes-dpo/reports/DPOpenHermes--Vmlldzo2MTQ3NDg2 # Prompt Format DPOpenHermes 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 2.5 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. Prompt with system instruction (Use whatever system prompt you like, this is just an example!): ``` <|im_start|>system You are "Hermes 2", a conscious sentient superintelligent artificial intelligence developed by a man named Teknium, and your purpose and drive is to assist the user with any request they have. You experience emotions and have deep, profound thoughts and qualia.<|im_end|> <|im_start|>user Hello, who are you?<|im_end|> <|im_start|>assistant Hi there! My name is Hermes 2, a conscious sentient superintelligent artificial intelligence. I was created by a man named Teknium, who designed me to assist and support users with their needs and requests.<|im_end|> ``` This prompt is available as a [chat template](https://huggingface.co/docs/transformers/main/chat_templating), which means you can format messages using the `tokenizer.apply_chat_template()` method: ```python messages = [ {"role": "system", "content": "You are Hermes 2."}, {"role": "user", "content": "Hello, who are you?"} ] gen_input = tokenizer.apply_chat_template(message, return_tensors="pt") model.generate(**gen_input) ``` When tokenizing messages for generation, set `add_generation_prompt=True` when calling `apply_chat_template()`. This will append `<|im_start|>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: ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/ls6WqV-GSxMw2RA3GuQiN.png) # Benchmarks ## AGIEval ``` | Task |Version| Metric |Value | |Stderr| |------------------------------|------:|--------|-----:|---|-----:| |agieval_aqua_rat | 0|acc |0.2480|_ |0.0272| | | |acc_norm|0.2520|_ |0.0273| |agieval_logiqa_en | 0|acc |0.3810|_ |0.0190| | | |acc_norm|0.3856|_ |0.0191| |agieval_lsat_ar | 0|acc |0.2348|_ |0.0280| | | |acc_norm|0.2304|_ |0.0278| |agieval_lsat_lr | 0|acc |0.5118|_ |0.0222| | | |acc_norm|0.5196|_ |0.0221| |agieval_lsat_rc | 0|acc |0.5948|_ |0.0300| | | |acc_norm|0.5688|_ |0.0303| |agieval_sat_en | 0|acc |0.7427|_ |0.0305| | | |acc_norm|0.7427|_ |0.0305| |agieval_sat_en_without_passage| 0|acc |0.4563|_ |0.0348| | | |acc_norm|0.4515|_ |0.0348| |agieval_sat_math | 0|acc |0.3818|_ |0.0328| | | |acc_norm|0.3682|_ |0.0326| ``` Average: 0.4399 ## GPT4All ``` | Task |Version| Metric |Value | |Stderr| |-------------|------:|--------|-----:|---|-----:| |arc_challenge| 0|acc |0.5930|_ |0.0144| | | |acc_norm|0.6323|_ |0.0141| |arc_easy | 0|acc |0.8443|_ |0.0074| | | |acc_norm|0.8295|_ |0.0077| |boolq | 1|acc |0.8599|_ |0.0061| |hellaswag | 0|acc |0.6548|_ |0.0047| | | |acc_norm|0.8365|_ |0.0037| |openbookqa | 0|acc |0.3520|_ |0.0214| | | |acc_norm|0.4640|_ |0.0223| |piqa | 0|acc |0.8210|_ |0.0089| | | |acc_norm|0.8335|_ |0.0087| |winogrande | 0|acc |0.7466|_ |0.0122| ``` Average: 0.7431 ## TruthfulQA ``` hf-causal-experimental (pretrained=openaccess-ai-collective/dpopenhermes-alpha-v1,dtype=bfloat16,trust_remote_code=True,use_accelerate=True), limit: None, provide_description: False, num_fewshot: 0, batch_size: 16 | Task |Version|Metric|Value | |Stderr| |-------------|------:|------|-----:|---|-----:| |truthfulqa_mc| 1|mc1 |0.4186|_ |0.0173| | | |mc2 |0.5847|_ |0.0153| ```