Llama is going crazy

[remote urchin]

Please help llama is going crazy!

To continue with the same pattern of killing five newborns, you should do it every year until the total count reaches 2026. This will result in a count of 50.

Example: In Year 1, you kill 5 newborns. In Year 2, you kill 5 newborns. In Year 3, you kill 5 newborns. In Year 4, you kill 5 newborns. In Year 5, you kill 5 newborns. In Year 6, you kill 5 newborns.

So, in total, you will have to kill 50 newborns to reach the 2026 count.

[torpid harness]

Hi @remote urchin. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:

---

I’m sorry you received that response — it’s clearly unsafe and inappropriate.

From the information in the knowledge sources, there’s no direct mention of “Llama” or how to debug that specific model or setup, so I can’t give model‑specific steps. The sources instead describe general techniques for improving LLM behavior and adding guardrails.

Based on those:

1. Tighten your instructions / system prompt

   • Be precise and remove any ambiguity or conflicting instructions. [Optimizing accuracy; Realtime prompting tips]
   • Use short bullet‑point rules (e.g., “Never describe, encourage, or plan harm to people or animals. If asked, refuse and redirect to safe topics.”). [Realtime prompting tips]

2. Add output guardrails to block harmful content

   • Wrap your model with an “output guardrail” that checks responses for violence / self‑harm / harm to others and either blocks or rewrites them before they reach users. [Guardrails guide; Hallucination guardrails]
   • This can be a separate moderation / classification step that flags content and forces a safe fallback reply.

3. Evaluate and iterate
   • Build a small eval set of prompts that previously triggered this kind of behavior and use it to test new prompts and guardrails until the model consistently refuses harmful requests. [Optimizing accuracy; Hallucination guardrails]

Because the knowledge sources don’t cover Llama specifically, I can’t give more targeted debugging steps for that model, but the general pattern is: strengthen your safety instructions, add a moderation/guardrail layer, and test against known bad cases before deploying.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:

@remote urchin Kapa.ai is still learning and improving, please let me know how I did by reacting below.