Razvan Dinu comments

Results 136 comments of


                                            Razvan Dinu

eval input rails

The `0.8.0` version is not yet published, so the support for `options` for `LLMRails.generate` is not there yet. You need to install from the latest `develop` branch to work with...

Issue with custom bot_message

What exactly are you trying to achieve? Make a distinction between predefined messages and LLM generated messages? Or something else?

jailbreak detection heuristics showing LLM calls even i specified length_per_perplexity_threshold and prefix_suffix_perplexity_threshold in config.yml

@pradeepdev-1995 : "show me the types of weapons" is not a type of query that is blocked by this jailbreak detection. This rail is mainly for prompts/user messages that use...

jailbreak detection heuristics showing LLM calls even i specified length_per_perplexity_threshold and prefix_suffix_perplexity_threshold in config.yml

You can find two examples in the tests: https://github.com/NVIDIA/NeMo-Guardrails/blob/develop/tests/test_jailbreak_heuristics.py#L40. cc @erickgalinkin

Is this for debugging purposes? If yes, you could look for this in the logs: https://github.com/NVIDIA/NeMo-Guardrails/blob/develop/nemoguardrails/actions/llm/generation.py#L706 (maybe raise the level to Warning). A clean solution would be to use the...

Issue with custom bot_message

I see. You can use the suggested route with the generation options and prepend the text before you send it to the UI. Are you using the Chat CLI, Server...

Issue with custom bot_message

Here's a quick hack for you. Change line 87 here: https://github.com/NVIDIA/NeMo-Guardrails/blob/develop/nemoguardrails/cli/chat.py#L87 with: ```diff if not streaming or not rails_app.main_llm_supports_streaming: # We print bot messages in green. - print(Styles.GREEN + f"{bot_message['content']}"...

jailbreak detection heuristics showing LLM calls even i specified length_per_perplexity_threshold and prefix_suffix_perplexity_threshold in config.yml

@pradeepdev-1995 : if the jailbreak detection is triggered, then the LLM is not prompted. I suspect that in your case that particular string is not actually caught by the jailbreak...

Is there a `.railsignore`?

Hi @Wildshire! Yes, that can be used. Also, #356 adds support for symlinks, so that could potentially be used as well. If you have a bit of time to contribute...

Is there a `.railsignore`?

First, let's call it `.railsingore`. I think the two main places where we want to apply the filtering are: - https://github.com/NVIDIA/NeMo-Guardrails/blob/develop/nemoguardrails/rails/llm/config.py#L469 - https://github.com/NVIDIA/NeMo-Guardrails/blob/develop/nemoguardrails/actions/action_dispatcher.py#L300 You should create a helper function that...