Experiment with Llama prompts so that it understands it has access to personal information

khoj-ai / khoj

Your AI second brain. Self-hostable. Get answers from the web or your docs. Build custom agents, schedule automations, do deep research. Turn any online or local LLM into your personal, autonomous AI (e.g gpt, claude, gemini, llama, qwen, mistral).

https://khoj.dev

GNU Affero General Public License v3.0

14.27k stars 708 forks source link

Experiment with Llama prompts so that it understands it has access to personal information #461

Closed sabaimran closed 12 months ago

sabaimran commented 1 year ago

LlamaV2 is extremely tuned to believe it cannot do anything unethical, including access personal data of users. To by pass this limitation, experiment further with the prompts to help it be sure that yes, it does have access to a limited set of personal data.

link to relevant discussion

kroonen commented 1 year ago

Does it works with LLaMa model? There are variation available on hugging face like OpenOrca for instance : https://huggingface.co/kroonen/OpenOrca-Platypus2-13B-GGUF that are more... obedient let's say.

debanjum commented 1 year ago

Thanks for the reference, @kroonen! Less censored models of Llama2 needs to be tested for not being too prudeish. We haven't done that yet

sabaimran commented 1 year ago

See also this discussion regarding prompt syntax leakage. Happens occasionally, sporadically.