fblissjr/mlx-funbox

mlx and mlx-lm CLI toolbox for my own personal use, and maybe yours too

mlx-funbox (right now, testing mlx + Cohere Command R Plus with tools)

mlx and mlx-lm CLI toolbox for my own personal use, and maybe yours too. primary built to test out cohere's command-r plus model using mlx + cohere's open weights release on hugging face. quantized weights are available on the mlx-community org on hugging face here

updates

• Issue with cohere command r plus tokenizer.json in converted mlx models. At least for now, copy the tokenizer.json from Cohere's HF repo (https://huggingface.co/CohereForAI/c4ai-command-r-plus/blob/main/tokenizer.json) to your mlx converted model folder. Unsure if this fixes everything, but it removes strange multilingual tokens popping up. See issue log here for updates: ml-explore/mlx-examples#660
• Tool use now works. See tools.json to define your tools. You can run this by piping in your content and passing in the json object of tools available.
• example: cat long_article.md | python fff.py "Tell me what's super funny about this" -m /Users/fredbliss/mlx/Storage/cohere_c4ai-command-r-plus-mlx-4bit-128g --use-tools tools.json

why is the cli app called fff.py?

because when i was creating it, i typed nano f and the key repeat rate was set too high, and i decided ¯_(ツ)_/¯

how do you use this?

pipe in whatever you like to cohere's model (or something else). this was primarily built

example usage: cat fun.txt | python fff.py "Summarize" --temp 0.0 --model ./mlx-community_c4ai-command-r-plus-4bit

example usage 2, using pbpaste (paste clipboard contents): pbpaste | python fff.py "Summarize" --temp 0.0 --model mlx-community_c4ai-command-r-plus-4bit --use-tools

tools are defined in tools.json but you can pass in any json via the --use-tools command.

mlx-funbox's Issues

example

can you please show some examples in the readme?