llm-gpt4all

Plugin for LLM adding support for the GPT4All collection of models.

Installation

Install this plugin in the same environment as LLM.

llm install llm-gpt4all

After installing the plugin you can see a new list of available models like this:

llm models list

The output will include something like this:

gpt4all: all-MiniLM-L6-v2-f16 - SBert, 43.76MB download, needs 1GB RAM (installed)
gpt4all: orca-mini-3b-gguf2-q4_0 - Mini Orca (Small), 1.84GB download, needs 4GB RAM (installed)
gpt4all: mistral-7b-instruct-v0 - Mistral Instruct, 3.83GB download, needs 8GB RAM (installed)
gpt4all: mistral-7b-openorca - Mistral OpenOrca, 3.83GB download, needs 8GB RAM (installed)
gpt4all: Meta-Llama-3-8B-Instruct - Llama 3 Instruct, 4.34GB download, needs 8GB RAM (installed) (aliases: l3)
gpt4all: all-MiniLM-L6-v2 - SBert, 43.82MB download, needs 1GB RAM
gpt4all: nomic-embed-text-v1 - Nomic Embed Text v1, 261.58MB download, needs 1GB RAM
gpt4all: nomic-embed-text-v1 - Nomic Embed Text v1.5, 261.58MB download, needs 1GB RAM
gpt4all: replit-code-v1_5-3b-newbpe-q4_0 - Replit, 1.82GB download, needs 4GB RAM
gpt4all: mpt-7b-chat - MPT Chat, 3.54GB download, needs 8GB RAM
gpt4all: orca-2-7b - Orca 2 (Medium), 3.56GB download, needs 8GB RAM
gpt4all: rift-coder-v0-7b-q4_0 - Rift coder, 3.56GB download, needs 8GB RAM
gpt4all: mpt-7b-chat-newbpe-q4_0 - MPT Chat, 3.64GB download, needs 8GB RAM
gpt4all: em_german_mistral_v01 - EM German Mistral, 3.83GB download, needs 8GB RAM
gpt4all: ghost-7b-v0 - Ghost 7B v0.9.1, 3.83GB download, needs 8GB RAM
gpt4all: Nous-Hermes-2-Mistral-7B-DPO - Nous Hermes 2 Mistral DPO, 3.83GB download, needs 8GB RAM
gpt4all: gpt4all-falcon-newbpe-q4_0 - GPT4All Falcon, 3.92GB download, needs 8GB RAM
gpt4all: gpt4all-13b-snoozy-q4_0 - Snoozy, 6.86GB download, needs 16GB RAM
gpt4all: wizardlm-13b-v1 - Wizard v1.2, 6.86GB download, needs 16GB RAM
gpt4all: orca-2-13b - Orca 2 (Full), 6.86GB download, needs 16GB RAM
gpt4all: nous-hermes-llama2-13b - Hermes, 6.86GB download, needs 16GB RAM
gpt4all: starcoder-newbpe-q4_0 - Starcoder, 8.37GB download, needs 4GB RAM

Further details on these models can be found in this Observable notebook.

Usage

Model Installation and Prompting

You can execute a model using the name displayed in the llm models list output. The model file will be downloaded the first time you attempt to run it.

llm -m orca-mini-3b-gguf2-q4_0 '3 names for a pet cow'

The first time you run this you will see a progress bar for the model file download:

 31%|█████████▋                        | 1.16G/3.79G [00:26<01:02, 42.0MiB/s]

On subsequent uses the model output will be displayed immediately.

Note that the models will be downloaded to ~/.cache/gpt4all.

Model options

Run llm models --options for a list of available model options, which should include:

gpt4all: mistral-7b-instruct-v0 - Mistral Instruct, 3.83GB download, needs 8GB RAM (installed)
  max_tokens: int
    The maximum number of tokens to generate.
  temp: float
    The model temperature. Larger values increase creativity but decrease
    factuality.
  top_k: int
    Randomly sample from the top_k most likely tokens at each generation
    step. Set this to 1 for greedy decoding.
  top_p: float
    Randomly sample at each generation step from the top most likely
    tokens whose probabilities add up to top_p.
  repeat_penalty: float
    Penalize the model for repetition. Higher values result in less
    repetition.
  repeat_last_n: int
    How far in the models generation history to apply the repeat penalty.
  n_batch: int
    Number of prompt tokens processed in parallel. Larger values decrease
    latency but increase resource requirements.

Use them like this:

llm -m mistral-7b-instruct-v0 -o max_tokens 2 'hi'

Hello!

Chatting

To chat with a model, avoiding the need to load it into memory for every message, use llm chat:

llm chat -m orca-mini-3b-gguf2-q4_0

Chatting with orca-mini-3b-gguf2-q4_0
Type 'exit' or 'quit' to exit
Type '!multi' to enter multiple lines, then '!end' to finish
> hi
 Hello! How can I assist you today?
> three jokes about a pelican and a chicken who are friends
 Sure, here are three jokes about a pelican and a chicken who are friends:

1. Why did the pelican cross the road? To get to the other side where the chicken was waiting for him!
2. What do you call a group of chickens playing basketball? A flock of feathers!
3. Why did the chicken invite the pelican over for dinner? Because it had nothing else to eat and needed some extra protein in its diet!

Removing Models

To remove a downloaded model, delete the .gguf file from ~/.cache/gpt4all.

Development

To set up this plugin locally, first checkout the code. Then create a new virtual environment:

cd llm-gpt4all
python3 -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

pip install -e '.[test]'

To run the tests:

pytest

simonw/llm-gpt4all