Bash vs shell code container

[beliashou]

Good day!

All my LLM when generate response used bash , not shell. And gptme don't execute the commands. What Do I wrong? How can I fix it?

[ErikBjare]

You are probably using a model that is bad at instruction following. If it cannot follow the simple system instructions, I wouldn't expect the commands to be very correct either.

You can try the --tool-format tool (might not be supported for your provider) or --tool-format xml to try other tool calling formats, but you should probably just use a better model (Claude is best, GPT-4o is good, Llama 405B should also work ok).

You can see some basic eval results for different models here: https://gptme.org/docs/evals.html

[beliashou]

I try to use ollama with deepseek-r1. It's not bad. Where can I see the system prompt for shell plugin?

[ErikBjare]

You can see the system prompt by running gptme with --show-hidden, or by reading the docs: https://gptme.org/docs/prompts.html

Deepseek R1 distillations in Ollama have been very disappointing for me when it comes to tool use/instruction following. Going to try a 70B distillation now that it's available on OpenRouter.

[ErikBjare]

I can confirm that https://openrouter.ai/deepseek/deepseek-r1-distill-llama-70b works pretty well