<img src=https://huggingface.co/lodrick-the-lafted/Hermes-Instruct-7B-217K/resolve/main/hermes-instruct.png>

Hermes-Instruct-7B-217K

Mistral-7B-Instruct-v0.2 trained with 217K rows of teknium/openhermes, in Alpaca format. Why? Mistral-7B-Instruct-v0.2 has native 32K context and rope theta of 1M. It's not a base model, so I've used the same recipe with different amounts of data to gauge the effects of further finetuning.

Prompt Format

Both the default Mistral-Instruct tags and Alpaca are fine, so either:

<s>[INST] {sys_prompt} {instruction} [/INST] 

or

{sys_prompt}

### Instruction:
{instruction}

### Response:

The tokenizer default is Alpaca this time around.

Usage

from transformers import AutoTokenizer
import transformers
import torch

model = "lodrick-the-lafted/Hermes-Instruct-7B-217K"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    model_kwargs={"torch_dtype": torch.bfloat16},
)

messages = [{"role": "user", "content": "Give me a cooking recipe for an apple pie."}]
prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.95)
print(outputs[0]["generated_text"])