UPD: Занавес. Оказалось, что reflection – кое-как зафайнтюненная llama 3.0 и даже не 3.1 и нормальных подвтерждений, что она обходит закрытые модели нет 😵‍💫

UPD 2: А Reflection API – это вообще Claude 3,5 Sonnet 😂 Подробности тут – https://www.reddit.com/r/LocalLLaMA/comments/1fc98fu/confirmed_reflection_70bs_official_api_is_sonnet/

Keep prompts simple and direct: The models excel at understanding and responding to brief, clear instructions without the need for extensive guidance.

Avoid chain-of-thought prompts: Since these models perform reasoning internally, prompting them to "think step by step" or "explain your reasoning" is unnecessary.

Use delimiters for clarity: Use delimiters like triple quotation marks, XML tags, or section noscripts to clearly indicate distinct parts of the input, helping the model interpret different sections appropriately.

Limit additional context in retrieval-augmented generation (RAG): When providing additional context or documents, include only the most relevant information to prevent the model from overcomplicating its response.