---
base_model: teknium/OpenHermes-2.5-Mistral-7B
tags:
- mistral
- instruct
- finetune
- chatml
- gpt4
- synthetic data
- distillation
- dpo
- rlhf
license: apache-2.0
language:
- en
datasets:
- mlabonne/chatml_dpo_pairs
---

> Note: This repository contains the `GGUF` 4-bit quantized variant of `halbihn/NeuralHermes-2.5-Mistral-7B` for the full version visit [the link](https://huggingface.co/halbihn/NeuralHermes-2.5-Mistral-7B)

<center><img src="https://i.imgur.com/qIhaFNM.png"></center>

# NeuralHermes 2.5 - Mistral 7B

NeuralHermes is based on the [teknium/OpenHermes-2.5-Mistral-7B](https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B) model that has been further fine-tuned with Direct Preference Optimization (DPO) using the [mlabonne/chatml_dpo_pairs](https://huggingface.co/datasets/mlabonne/chatml_dpo_pairs) dataset. It surpasses the original model on most benchmarks (see results).

It is directly inspired by the RLHF process described by [Intel/neural-chat-7b-v3-1](https://huggingface.co/Intel/neural-chat-7b-v3-1)'s authors to improve performance. I used the same dataset and reformatted it to apply the ChatML template.

The code to train this model is available on [Google Colab](https://colab.research.google.com/drive/1h4tAJStIef_BcO-OkY97X9_OFgKnFrLl). It required an A100 GPU for about an hour.

## Quantized models

* **GGUF**: https://huggingface.co/TheBloke/NeuralHermes-2.5-Mistral-7B-GGUF
* **AWQ**: https://huggingface.co/TheBloke/NeuralHermes-2.5-Mistral-7B-AWQ
* **GPTQ**: https://huggingface.co/TheBloke/NeuralHermes-2.5-Mistral-7B-GPTQ
* **EXL2**:
  * 3.0bpw: https://huggingface.co/LoneStriker/NeuralHermes-2.5-Mistral-7B-3.0bpw-h6-exl2
  * 4.0bpw: https://huggingface.co/LoneStriker/NeuralHermes-2.5-Mistral-7B-4.0bpw-h6-exl2
  * 5.0bpw: https://huggingface.co/LoneStriker/NeuralHermes-2.5-Mistral-7B-5.0bpw-h6-exl2
  * 6.0bpw: https://huggingface.co/LoneStriker/NeuralHermes-2.5-Mistral-7B-6.0bpw-h6-exl2
  * 8.0bpw: https://huggingface.co/LoneStriker/NeuralHermes-2.5-Mistral-7B-8.0bpw-h8-exl2

## Results

**Update:** NeuralHermes-2.5 became the best Hermes-based model on the Open LLM leaderboard and one of the very best 7b models. 🎉

![image/png](https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/yWe6VBFxkHiuOlDVBXtGo.png)

Teknium (author of OpenHermes-2.5-Mistral-7B) benchmarked the model ([see his tweet](https://twitter.com/Teknium1/status/1729955709377503660)).

Results are improved on every benchmark: **AGIEval** (from 43.07% to 43.62%), **GPT4All** (from 73.12% to 73.25%), and **TruthfulQA**.

### AGIEval
![](https://i.imgur.com/7an3B1f.png)

### GPT4All
![](https://i.imgur.com/TLxZFi9.png)

### TruthfulQA
![](https://i.imgur.com/V380MqD.png)

You can view the Weights & Biases report [here](https://api.wandb.ai/links/halbihn/uem1q2dj).

## Usage

You can run this model using [LM Studio](https://lmstudio.ai/) or any other frontend.

You can also run this model using the following code:

```python
import transformers
from transformers import AutoTokenizer

model_id = "halbihn/NeuralHermes-2.5-Mistral-7B"

# Format prompt
message = [
    {"role": "system", "content": "You are a helpful assistant chatbot."},
    {"role": "user", "content": "What is a Large Language Model?"}
]
tokenizer = AutoTokenizer.from_pretrained(model_id)
prompt = tokenizer.apply_chat_template(message, add_generation_prompt=True, tokenize=False)

# Create pipeline
pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    tokenizer=tokenizer
)

# Generate text
sequences = pipeline(
    prompt,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    num_return_sequences=1,
    max_length=200,
)
response = sequences[0]['generated_text'].split("<|im_start|>assistant")[-1].strip()
print(response)


# streaming example
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
import torch

model_id = "halbihn/NeuralHermes-2.5-Mistral-7B"

model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model.to(device)

def stream(
  user_prompt: str,
  max_tokens: int = 200,
) -> None:
  """Text streaming example
  """
  
  system_prompt = 'Below is a conversation between Human and AI assistant named Mistral\n'

  message = [
      {"role": "system", "content": system_prompt},
      {"role": "user", "content": user_prompt}
  ]
  prompt = tokenizer.apply_chat_template(
    message, 
    add_generation_prompt=True, 
    tokenize=False,
  )

  inputs = tokenizer([prompt], return_tensors="pt").to(device)

  streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

  _ = model.generate(**inputs, streamer=streamer, max_new_tokens=max_tokens)

stream("Tell me about the future")

>>> The future is a vast and uncertain expanse, shaped by the collective actions and innovations of humanity. It is a blend of possibilities, technological advancements, and societal changes. Some potential aspects of the future include:
>>> 
>>> 1. Technological advancements: Artificial intelligence, quantum computing, and biotechnology are expected to continue evolving, leading to breakthroughs in fields like medicine, energy, and communication.
>>> 
>>> 2. Space exploration: As technology progresses, space travel may become more accessible, enabling humans to establish colonies on other planets and explore the cosmos further.
>>> 
>>> 3. Climate change mitigation: The future will likely see increased efforts to combat climate change through renewable energy sources, carbon capture technologies, and sustainable practices.
>>> 
>>> 4. Artificial intelligence integration: AI will likely become more integrated into daily life, assisting with tasks, automating jobs, and even influencing decision-making processes in various industries.
```

## Training hyperparameters

**LoRA**:
* r=16
* lora_alpha=16
* lora_dropout=0.05
* bias="none"
* task_type="CAUSAL_LM"
* target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']

**Training arguments**:
* per_device_train_batch_size=4
* gradient_accumulation_steps=4
* gradient_checkpointing=True
* learning_rate=5e-5
* lr_scheduler_type="cosine"
* max_steps=200
* optim="paged_adamw_32bit"
* warmup_steps=100

**DPOTrainer**:
* beta=0.1
* max_prompt_length=1024
* max_length=1536