IHOP (Instruct Humanoid Optimising Protocol)

IHOP is a tool for optimizing and compiling opensource instruct datasets to create better finetunings for large language models.

The IHOP format

Messages are sent to the model in the following format:

<|SYSTEM|>You are a helpful executive assitant. Your Job is to respond to user inputs with helpful and factual answers. The user may pass context, if context is provided use the context combined with your knowledge to provide answers.<|END_SYSTEM|>  # The system message defaults to 
<|CONTEXT|>News articles or websites, etc<|CONTEXT|>  # Context is optional
<|USER_INPUT|>The user input<|END_USER_INPUT|>
<|RESPONSE|>  # always end in the opening response tag

User <|USER_INPUT|> and <|RESPONSE|> tags can be chained to continue chat converstations.

Running server.py with sanic server --workers=1 workers can be increased to increase throughput, but each worker keeps an instance of the model in vram. The sanic server is also compatible with OpenAI Chat Completions endpoint with added support for the context message. Full support is coming in the future. An example api request can be seen here https://github.com/getorca/IHOP/blob/main/scripts/test_chat_api.sh. The API response is raw text of only the response.

IHOP Models

Model	Base Model	Dataset	Description	Training
IHOPv01_Falcon7b_lora	Falcon 7b	IHOPv01	An experimental general purpose instruct following chat model	This repo
pb_lora_7b_v0.1	llama 7b	reddit_finance_43_250k	An experimental model trained to reply to various finance, crypto and investing subreddits	Training

IHOP Datasets

Dataset	Size	Source	Description	Recreation
IHOPv01	67.4k	composite	A collecton of instruct datasets, curated to not reject instructions	Building

Run training

Run the training with accelerate and deepspeed https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#deepspeed

1 - run accelerate config answer the options for your training system 2 - run training with accelerate launch lora_falcon_finetune.py --output_dir loras/IHOP_lora_falcon_7b_v1