Used to train models that interact directly with shells

Note: This dataset is out-dated in the llm world, probably easier to just setup a tool with a decent model that supports tooling.

Follow-up details of my process

• MacOS terminal commands for now. This dataset is still in alpha stages and will be modified.

• Contains 500 somewhat unique training examples so far.

• GPT4 seems like a good candidate for generating more data, licensing would need to be addressed.

• I fine-tuned Solar-10.7B-Instruct-v1.0 with this dataset using a slightly modified version of axolotl. Just a few epochs was enough to get it to output correctly.

• I use oobabooga/text-generation-webui with a custom chat extension for inference. No sandbox is used, it is piped directly into MacOS bash because I'm reckless. C:

• Currently working towards training an MoE (2x7B), multi-modal model (image/text) with this dataset. (BakLLaVA-1-7B + LLaVA-v1.5-7B)

• Inference stages:

  1. Send the instruction to the model, expect command.
  2. Detect shell command and send to sand-boxed shell.
  3. Shell respose should be sent as additional input to model.
  4. The final model response should be sent to user from assistant.

TODO:

• Possible "os" column to specify which system the command should be used with, maybe separate datasets for each system type.

Sample prompt: (in series, depends on your specific model prompt)

### User:
List files in 'Downloads'

### Command:	
ls ~/Downloads

### Shell:
file1.pdf file2.txt file3.zip

### Assistant:	
Listing files in 'Downloads': file1.pdf file2.txt file3.zip