Senior software engineer at Qualia Labs · Co-founder of Fox.Build Makerspace · Former co-founder of FarmBot

Promptingguide.ai

Re-visiting PromptingGuide.AI since it had a bunch of updates since the last time I read it.

Frequency Penalty - Applies a penalty to frequently repeated tokens in the response and prompt, reducing the likelihood of repeated words by assigning higher penalties to tokens that appear more frequently.

Presence Penalty - Applies an equal penalty to all repeated tokens, regardless of frequency, which helps in generating more diverse responses by discouraging repeated phrases without regard to repetition count.

Top P - A sampling technique that limits responses to tokens within a probability threshold, with lower values favoring confident responses and higher values increasing diversity by including less likely options.

Zero-shot prompting - A technique where the model performs a task without examples, relying on general language knowledge to interpret and respond appropriately.

Few-shot prompting - Similar to zero-shot prompting, but includes examples to guide the model’s response style or format.

Few-shot pro-tip - Selecting random labels from a true distribution of labels (instead of a uniform distribution) also helps.

Interesting tidbit from the part on CoT prompting:

...the authors claim that this is an emergent ability that arises with sufficiently large language models.

Meta-prompting - Similar to CoT prompting but the focus is on the structure of the prompt language rather than examples.

Prompt Chaining - Common use of multi-step prompts. Anthropic docs.

  • [ ] Try a tree of thought prompt. Docs were not deep enough.
  • [ ] What is Named entity recognition?
  • [ ] Find some real world uses of ReAct prompting
    • [ ] Look into AlfWorld ReAct examples
  • [ ] Read the papers presented in the page about data generation. Use GPT to summarize if needed. To tackle the diversity issue, the authors prepared a vocabulary of around 1500 basic words, mirroring a typical child's vocabulary, divided into nouns, verbs, and adjectives. In each generation, one verb, one noun, and one adjective were randomly selected. The model then generates a story integrating these random words.
Write a short story (3-5 paragraphs) which only uses very simple words that a 3 year old child would likely understand. The story should use the verb ”{random.choice(verbs_list)}”, the noun ”{random.choice(nouns_list)}” and the adjective ”{random.choice(adjectives_list)}”. The story should have the following features: {random.choice(features_list)}, {random.choice(features_list)}. Remember to only use simple words!
  • [x] Try this for sentence generation :point_up:
주제: {주제 입력}

말투: {말투 입력} (예: 격식체, 비격식체, 구어체, 문어체 등)

위의 주제에 대해 {말투}로 한 단락의 짧은 글을 작성해주세요. 이 글은 한국어 학습자를 위한 읽기 자료입니다.
{
    "환경 보전": "격식체",
    "인공지능": "비격식체",
    "우정": "구어체",
    "우주 탐사": "문어체",
    "건강한 식습관": "일상체",
    "문화적 다양성": "격식체",
    "교육에서의 기술": "비격식체",
    "기후 변화": "과학체",
    "해외 여행": "비격식체",
    "독서의 중요성": "문어체",
    "운동과 체력": "일상체",
    "사이버 보안": "격식체",
    "마음챙김과 명상": "비격식체",
    "소셜 미디어의 영향": "구어체",
    "예술의 역사": "문어체",
    "자원 봉사": "격식체",
    "디지털 노마드 라이프스타일": "일상체",
    "재무 지식": "비격식체",
    "리더십 기술": "격식체",
    "도시화": "과학체"
}

Experimental prompt output