Llama Prompts — Meta AI & Open-Source LLM Tips

I'm setting up Llama {{model_version}} locally on a machine with {{gpu_specs}}. Help me create the optimal deployment configuration.

Cover:
1. Recommended quantization level for my hardware (Q4_K_M, Q5_K_M, Q8_0, or FP16) with reasoning
2. The exact ollama/llama.cpp command to download and run the model
3. Optimal context window size for my VRAM
4. Recommended inference parameters (temperature, top_p, top_k, repeat_penalty) for {{use_case}}
5. Expected tokens/second performance range
6. Memory usage estimate (RAM + VRAM)

If my hardware can't run this model well, suggest the largest model variant that would work smoothly.

I want to fine-tune Llama {{model_version}} for {{task_description}}. Help me prepare a training dataset.

Generate {{num_examples}} high-quality training examples in the chat format Llama expects:

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

Requirements:
- System prompt should be consistent across all examples and optimized for {{task_description}}
- User messages should vary in complexity (simple, moderate, complex)
- Assistant responses should demonstrate the exact style, format, and quality I want
- Include edge cases the model might encounter in production
- Each example should teach the model something slightly different

Also provide:
- Recommended number of total training examples for this task type
- Suggested LoRA hyperparameters (rank, alpha, learning rate, epochs)
- How to validate the fine-tune worked

Compare quantization options for Llama {{model_version}} ({{parameter_count}} parameters) and help me choose the right one.

Create a comparison table with these quantization levels:
- Q4_K_M
- Q5_K_M
- Q6_K
- Q8_0
- FP16 (no quantization)

For each, show:
| Quant | File Size | VRAM Required | Quality Loss | Speed (tokens/s estimate) | Best For |

My hardware: {{hardware_specs}}
My use case: {{use_case}}
My priority: {{priority}} (quality / speed / memory efficiency)

Recommend the best option for my setup and explain what quality tradeoffs I'm making. Include the exact command to convert or download at that quantization level.

I'm running Llama {{model_version}} via {{inference_engine}} and getting {{current_speed}} tokens/second. I need to optimize for {{optimization_goal}}.

Current setup:
- Hardware: {{hardware_specs}}
- Quantization: {{current_quant}}
- Context length: {{context_length}}
- Batch size: {{batch_size}}

Provide a step-by-step optimization plan:
1. Quick wins (parameter changes, no code modifications)
2. Medium effort (configuration changes, engine swaps)
3. Advanced (custom kernels, batching strategies, speculative decoding)

For each suggestion, estimate the expected improvement and any tradeoffs. Prioritize by impact-to-effort ratio.

Design a RAG (Retrieval-Augmented Generation) pipeline using Llama {{model_version}} running locally.

My documents: {{document_description}}
My query types: {{query_types}}
Infrastructure: {{infrastructure}}

Provide:
1. **Embedding model recommendation** — which model to use for embeddings (local options), with dimension size and performance notes
2. **Chunking strategy** — optimal chunk size, overlap, and splitting method for my document types
3. **Vector store setup** — recommended local vector DB (ChromaDB, Qdrant, FAISS) with config
4. **Retrieval prompt template** — the exact system prompt and user prompt template that tells Llama to answer from retrieved context only
5. **Reranking** — whether to add a reranker and which one
6. **Evaluation** — how to measure if the RAG pipeline is working well

Include code snippets for the critical pieces (chunking, retrieval, prompt assembly).

Write a production-ready system prompt for Llama {{model_version}} that will be used as {{role_description}} in a {{application_type}} application.

The system prompt must:
- Define the model's role, personality, and boundaries clearly
- Specify the output format ({{output_format}})
- Include safety guardrails appropriate for {{deployment_context}}
- Handle edge cases: off-topic requests, harmful content, questions it can't answer
- Be optimized for Llama's chat template format

Also provide:
- A test suite of 5 user messages to verify the system prompt works (include adversarial examples)
- The expected response behavior for each test message
- Recommended temperature and sampling parameters for this use case

Keep the system prompt under 500 tokens to maximize context window for user content.