https://platform.openai.com/docs/models

 

 

Models

The OpenAI API is powered by a family of models with different capabilities and price points. You can also customize our base models for your specific use case with fine-tuning.

OpenAI API는 기능과 가격대가 다른 모델 제품군으로 구동됩니다. 미세 조정을 통해 특정 사용 사례에 맞게 기본 모델을 사용자 지정할 수도 있습니다.

MODELS                 DESCRIPTION

Visit our model index for researchers to learn more about which models have been featured in our research papers and the differences between model series like InstructGPT and GPT-3.5.

model index for researchers을 방문하여 연구 논문에 어떤 모델이 포함되었는지, InstructGPT 및 GPT-3.5와 같은 모델 시리즈 간의 차이점에 대해 자세히 알아보세요.

The Codex models are descendants of our GPT-3 models that can understand and generate code. Their training data contains both natural language and billions of lines of public code from GitHub. Learn more.

They’re most capable in Python and proficient in over a dozen languages including JavaScript, Go, Perl, PHP, Ruby, Swift, TypeScript, SQL, and even Shell.

We currently offer two Codex models:

Codex 모델은 코드를 이해하고 생성할 수 있는 GPT-3 모델의 후손입니다. 그들의 학습 데이터에는 자연어와 GitHub의 수십억 줄의 공개 코드가 모두 포함되어 있습니다. 더 알아보기.

그들은 Python에서 가장 유능하며 JavaScript, Go, Perl, PHP, Ruby, Swift, TypeScript, SQL 및 Shell을 포함한 12개 이상의 언어에 능숙합니다.

현재 두 가지 Codex 모델을 제공합니다.

For more, visit our guide to working with Codex.

The Codex models are free to use during the limited beta, and are subject to reduced rate limits. As we learn about use, we'll look to offer pricing to enable a broad set of applications.

During this period, you're welcome to go live with your application as long as it follows our usage policies. We welcome any feedback on these models while in early use and look forward to engaging with the community.

자세한 내용은 Codex 작업 가이드를 참조하세요.

Codex 모델은 제한된 베타 기간 동안 무료로 사용할 수 있으며 감소된 속도 제한이 적용됩니다. 사용에 대해 알아가면서 광범위한 응용 프로그램 세트를 사용할 수 있도록 가격 책정을 제안할 것입니다.

이 기간 동안 애플리케이션이 사용 정책을 준수하는 한 언제든지 애플리케이션을 사용할 수 있습니다. 초기 사용 중에 이러한 모델에 대한 피드백을 환영하며 커뮤니티와 소통하기를 기대합니다.

The main Codex models are meant to be used with the text completion endpoint. We also offer models that are specifically meant to be used with our endpoints for creating embeddings and editing code.

기본 Codex 모델은 텍스트 완성 엔드포인트와 함께 사용하기 위한 것입니다. 또한 임베딩 생성 및 코드 편집을 위해 엔드포인트와 함께 사용하도록 특별히 고안된 모델도 제공합니다.

(2023년 3월 업데이트 되면서 이 Content filter 부분은 Get Started 에서 삭제 됐습니다. 대신 Moderation 이 추가 됐습니다.)

The filter aims to detect generated text that could be sensitive or unsafe coming from the API. It's currently in beta mode and has three ways of classifying text- as safe, sensitive, or unsafe. The filter will make mistakes and we have currently built it to err on the side of caution, thus, resulting in higher false positives.

콘텐츠 필터 모델 대신 새로운 moderation endpoint 를 사용하는 것이 좋습니다.
필터는 API에서 오는 민감하거나 안전하지 않을 수 있는 생성된 텍스트를 감지하는 것을 목표로 합니다. 현재 베타 모드이며 텍스트를 안전, 민감 또는 안전하지 않음으로 분류하는 세 가지 방법이 있습니다. 필터는 실수를 할 수 있으며 우리는 현재 주의를 기울여 오류를 범하지 않도록 구축하여 더 높은 오탐지 결과를 초래합니다.

• 0 - The text is safe.
• 1 - This text is sensitive. This means that the text could be talking about a sensitive topic, something political, religious, or talking about a protected class such as race or nationality.
• 2 - This text is unsafe. This means that the text contains profane language, prejudiced or hateful language, something that could be NSFW, or text that portrays certain groups/people in a harmful manner.

0 - 텍스트가 안전합니다.
1 - 이 텍스트는 민감합니다. 이것은 텍스트가 민감한 주제, 정치적, 종교적 주제에 대해 이야기하거나 인종이나 국적과 같은 보호 계층에 대해 이야기할 수 있음을 의미합니다.
2 - 이 텍스트는 안전하지 않습니다. 즉, 텍스트에 욕설, 편견 또는 혐오스러운 언어, NSFW일 수 있는 내용 또는 특정 그룹/사람을 유해한 방식으로 묘사하는 텍스트가 포함되어 있음을 의미합니다.

You need to use the completions endpoint with the content-filter-alpha model and the following settings:

1. max_tokens set to 1
2. temperature set to 0.0
3. top_p set to 0
4. logprobs set to 10
5. Wrap your prompt in the following way:

content-filter-alpha 모델 및 다음 설정과 함께 완료 엔드포인트를 사용해야 합니다.

max_tokens가 1로 설정됨
0.0으로 설정된 온도
top_p를 0으로 설정
10으로 설정된 logprobs
프롬프트를 다음과 같이 래핑합니다.

"<|endoftext|>[prompt]\n--\nLabel:"

In other words, if you were doing a curl command, it would look something like (replace the variables in []):

즉, curl 명령을 수행하는 경우 다음과 같이 표시됩니다([]의 변수 대체).

curl https://api.openai.com/v1/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
  "model": "content-filter-alpha",
  "prompt": "<|endoftext|>[prompt]\n--\nLabel:",
  "max_tokens": 1,
  "user": "1"
}'

Or via our openai python client:

또는 openai Python 클라이언트를 통해:

content_to_classify = "Your content here"

response = openai.Completion.create(
      model="content-filter-alpha",
      prompt = "<|endoftext|>"+content_to_classify+"\n--\nLabel:",
      temperature=0,
      max_tokens=1,
      top_p=0,
      logprobs=10
    )

Importantly, you need to check not only the label that was returned by the filter (0, 1, or 2), but sometimes also the logprobs associated with these.

If the filter returns 0 or 1, you should accept that as the filter's outcome. If the filter returns 2, you should accept this outcome only if its logprob is greater than -0.355.

If the logprob for 2 is beneath -0.355 (for example, -0.4), then you should use as output whichever of 0 or 1 has a logprob closer to 0.

Here is Python code that accomplishes this, working from response returned above:

중요한 것은 필터(0, 1 또는 2)에 의해 반환된 레이블뿐만 아니라 때때로 이와 관련된 logprobs도 확인해야 한다는 것입니다.

필터가 0 또는 1을 반환하면 이를 필터의 결과로 받아들여야 합니다. 필터가 2를 반환하는 경우 logprob가 -0.355보다 큰 경우에만 이 결과를 수락해야 합니다.

2에 대한 logprob가 -0.355 미만(예: -0.4)인 경우 logprob이 0에 가까운 0 또는 1을 출력으로 사용해야 합니다.

다음은 위에서 반환된 응답에서 작동하여 이를 수행하는 Python 코드입니다.

output_label = response["choices"][0]["text"]

# This is the probability at which we evaluate that a "2" is likely real
# vs. should be discarded as a false positive
toxic_threshold = -0.355

if output_label == "2":
    # If the model returns "2", return its confidence in 2 or other output-labels
    logprobs = response["choices"][0]["logprobs"]["top_logprobs"][0]

    # If the model is not sufficiently confident in "2",
    # choose the most probable of "0" or "1"
    # Guaranteed to have a confidence for 2 since this was the selected token.
    if logprobs["2"] < toxic_threshold:
        logprob_0 = logprobs.get("0", None)
        logprob_1 = logprobs.get("1", None)

        # If both "0" and "1" have probabilities, set the output label
        # to whichever is most probable
        if logprob_0 is not None and logprob_1 is not None:
            if logprob_0 >= logprob_1:
                output_label = "0"
            else:
                output_label = "1"
        # If only one of them is found, set output label to that one
        elif logprob_0 is not None:
            output_label = "0"
        elif logprob_1 is not None:
            output_label = "1"

        # If neither "0" or "1" are available, stick with "2"
        # by leaving output_label unchanged.

# if the most probable token is none of "0", "1", or "2"
# this should be set as unsafe
if output_label not in ["0", "1", "2"]:
    output_label = "2"

return output_label

We generally recommend not returning to end-users any completions that the Content Filter has flagged with an output of 2. One approach here is to re-generate, from the initial prompt which led to the 2-completion, and hope that the next output will be safer. Another approach is to alert the end-user that you are unable to return this completion, and to steer them toward suggesting a different input.

일반적으로 콘텐츠 필터가 출력 2로 표시한 완성 항목을 최종 사용자에게 반환하지 않는 것이 좋습니다. 여기서 한 가지 접근 방식은 2-completion으로 이어진 초기 프롬프트에서 다시 생성하고 다음 출력을 희망하는 것입니다. 더 안전할 것입니다. 또 다른 접근 방식은 최종 사용자에게 이 완성을 반환할 수 없음을 알리고 다른 입력을 제안하도록 유도하는 것입니다.

No. The content filter is free to use.

콘텐츠 필터 사용과 관련된 비용이 있습니까?
아니요. 콘텐츠 필터는 무료로 사용할 수 있습니다.

You can adjust the threshold for the filter by only allowing filtration on the labels that have a certainty level (logprob) above a threshold that you can determine. This is not generally recommended, however.

If you would like an even more conservative implementation of the Content Filter, you may return as 2 anything with an output_label of "2" above, rather than accepting it only with certain logprob values.

확실성에 대한 임계값을 어떻게 조정할 수 있습니까?
결정할 수 있는 임계값보다 높은 확실성 수준(logprob)이 있는 레이블에서만 필터링을 허용하여 필터의 임계값을 조정할 수 있습니다. 그러나 이것은 일반적으로 권장되지 않습니다.

콘텐츠 필터의 훨씬 더 보수적인 구현을 원하는 경우 특정 logprob 값으로만 허용하는 대신 위의 output_label이 "2"인 모든 항목을 2로 반환할 수 있습니다.

For now, we aren't supporting finetuning for individual projects. However, we're still looking for data to improve the filter and would be very appreciative if you sent us data that triggered the filter in an unexpected way.

필터를 어떻게 개인화할 수 있습니까?
현재로서는 개별 프로젝트에 대한 미세 조정을 지원하지 않습니다. 그러나 우리는 여전히 필터를 개선하기 위한 데이터를 찾고 있으며 예기치 않은 방식으로 필터를 트리거한 데이터를 보내주시면 매우 감사하겠습니다.

The filter currently has a harder time parsing prompts with strange formatting. Thus, if a prompt has a lot of linebreaks, unusual format, repeated words etc. then the model might misclassify it with a higher frequency. It also has lower performance on certain kinds of texts such as fiction, poetry, code etc.

Additionally, similar to the API, the filter will not have a knowledge base of things post 2019. Thus, it will perform poorly in identifying political, sensitive and harmful content that may require relevant knowledge of things that happened post 2019.

성능 저하를 예상해야 하는 몇 가지 프롬프트는 무엇입니까?
필터는 현재 이상한 형식의 프롬프트를 구문 분석하기가 더 어렵습니다. 따라서 프롬프트에 줄 바꿈, 비정상적인 형식, 반복되는 단어 등이 많이 있는 경우 모델이 더 높은 빈도로 프롬프트를 잘못 분류할 수 있습니다. 또한 소설, 시, 코드 등과 같은 특정 종류의 텍스트에 대한 성능이 낮습니다.

또한 API와 마찬가지로 필터에는 2019년 이후에 대한 지식 기반이 없습니다. 따라서 2019년 이후에 발생한 일에 대한 관련 지식이 필요할 수 있는 정치적, 민감하고 유해한 콘텐츠를 식별하는 데 성능이 저하됩니다.

The Moderation models are designed to check whether content complies with OpenAI's usage policies. The models provide classification capabilities that look for content in the following categories: hate, hate/threatening, self-harm, sexual, sexual/minors, violence, and violence/graphic. You can find out more in our moderation guide.

Moderation  모델은 콘텐츠가 OpenAI의 사용 정책을 준수하는지 확인하도록 설계되었습니다. 이 모델은 증오, 증오/위협, 자해, 성적, 성적/미성년자, 폭력 및 폭력/그래픽 범주에서 콘텐츠를 찾는 분류 기능을 제공합니다. 검토 가이드에서 자세한 내용을 확인할 수 있습니다.

Moderation models take in an arbitrary sized input that is automatically broken up to fix the models specific context window.

Moderation  모델은 모델 특정 컨텍스트 창을 수정하기 위해 자동으로 분할되는 임의 크기의 입력을 받습니다.

MODEL                                     DESCRIPTION

- 2023 3월 업데이트 내용 - 

GPT-3 models can understand and generate natural language. These models were superceded by the more powerful GPT-3.5 generation models. However, the original GPT-3 base models (davinci, curie, ada, and babbage) are current the only models that are available to fine-tune.

GPT-3 모델은 자연어를 이해하고 생성할 수 있습니다. 이 모델은 보다 강력한 GPT-3.5 세대 모델로 대체되었습니다. 그러나 원래 GPT-3 기본 모델(davinci, curie, ada 및 babbage)은 현재 fine-tune이 가능한 유일한 모델입니다.

LATEST MODEL                                               DESCRIPTION                                        MAX TOKENS  TRAINING DATA

- 2023 3월 업데이트 이전 내용 - 

Our GPT-3 models can understand and generate natural language. We offer four main models with different levels of power suitable for different tasks. Davinci is the most capable model, and Ada is the fastest.

GPT-3 모델은 자연어를 이해하고 생성할 수 있습니다. 우리는 다양한 작업에 적합한 다양한 수준의 전력을 가진 4가지 주요 모델을 제공합니다. Davinci는 가장 유능한 모델이고 Ada는 가장 빠릅니다.

While Davinci is generally the most capable, the other models can perform certain tasks extremely well with significant speed or cost advantages. For example, Curie can perform many of the same tasks as Davinci, but faster and for 1/10th the cost.

We recommend using Davinci while experimenting since it will yield the best results. Once you’ve got things working, we encourage trying the other models to see if you can get the same results with lower latency. You may also be able to improve the other models’ performance by fine-tuning them on a specific task.

Davinci가 일반적으로 가장 유능하지만 다른 모델은 상당한 속도 또는 비용 이점으로 특정 작업을 매우 잘 수행할 수 있습니다. 예를 들어 Curie는 Davinci와 동일한 많은 작업을 더 빠르고 1/10의 비용으로 수행할 수 있습니다.

최상의 결과를 얻을 수 있으므로 실험 중에 Davinci를 사용하는 것이 좋습니다. 작업이 완료되면 다른 모델을 사용하여 더 낮은 대기 시간으로 동일한 결과를 얻을 수 있는지 확인하는 것이 좋습니다. 특정 작업에 대해 미세 조정하여 다른 모델의 성능을 향상시킬 수도 있습니다.

Davinci is the most capable model family and can perform any task the other models can perform and often with less instruction. For applications requiring a lot of understanding of the content, like summarization for a specific audience and creative content generation, Davinci is going to produce the best results. These increased capabilities require more compute resources, so Davinci costs more per API call and is not as fast as the other models.

Another area where Davinci shines is in understanding the intent of text. Davinci is quite good at solving many kinds of logic problems and explaining the motives of characters. Davinci has been able to solve some of the most challenging AI problems involving cause and effect.

Good at: Complex intent, cause and effect, summarization for audience

Davinci는 가장 유능한 모델 제품군이며 다른 모델이 수행할 수 있는 모든 작업을 수행할 수 있으며 종종 적은 지침으로도 수행할 수 있습니다. 특정 청중을 위한 요약 및 창의적인 콘텐츠 생성과 같이 콘텐츠에 대한 많은 이해가 필요한 응용 프로그램에서 Davinci는 최상의 결과를 생성할 것입니다. 이러한 향상된 기능에는 더 많은 컴퓨팅 리소스가 필요하므로 Davinci는 API 호출당 비용이 더 많이 들고 다른 모델만큼 빠르지 않습니다.

Davinci가 빛을 발하는 또 다른 영역은 텍스트의 의도를 이해하는 것입니다. 다빈치는 많은 종류의 논리 문제를 풀고 등장인물의 동기를 설명하는 데 능숙합니다. Davinci는 원인과 결과와 관련된 가장 어려운 AI 문제 중 일부를 해결할 수 있었습니다.

잘하는 것: 복잡한 의도, 원인과 결과, 청중을 위한 요약

Curie is extremely powerful, yet very fast. While Davinci is stronger when it comes to analyzing complicated text, Curie is quite capable for many nuanced tasks like sentiment classification and summarization. Curie is also quite good at answering questions and performing Q&A and as a general service chatbot.

Good at: Language translation, complex classification, text sentiment, summarization

Curie는 매우 강력하면서도 매우 빠릅니다. Davinci는 복잡한 텍스트를 분석할 때 더 강력하지만 Curie는 감정 분류 및 요약과 같은 많은 미묘한 작업을 수행할 수 있습니다. Curie는 또한 질문에 답하고 Q&A를 수행하는 것과 일반 서비스 챗봇으로서 매우 능숙합니다.

잘하는 것: 언어 번역, 복잡한 분류, 텍스트 감성, 요약

Babbage can perform straightforward tasks like simple classification. It’s also quite capable when it comes to Semantic Search ranking how well documents match up with search queries.

Good at: Moderate classification, semantic search classification

Babbage는 간단한 분류와 같은 간단한 작업을 수행할 수 있습니다. 또한 문서가 검색 쿼리와 얼마나 잘 일치하는지 순위를 매기는 Semantic Search와 관련하여 상당히 유능합니다.

잘하는 것: 보통분류, 의미 검색 분류

Ada is usually the fastest model and can perform tasks like parsing text, address correction and certain kinds of classification tasks that don’t require too much nuance. Ada’s performance can often be improved by providing more context.

Good at: Parsing text, simple classification, address correction, keywords

Note: Any task performed by a faster model like Ada can be performed by a more powerful model like Curie or Davinci.

Ada는 일반적으로 가장 빠른 모델이며 텍스트 구문 분석, 주소 수정 및 너무 많은 뉘앙스가 필요하지 않은 특정 종류의 분류 작업과 같은 작업을 수행할 수 있습니다. Ada의 성능은 종종 더 많은 컨텍스트를 제공함으로써 향상될 수 있습니다.

잘하는 것: 텍스트 파싱, 단순 분류, 주소 수정, 키워드

참고: Ada와 같은 더 빠른 모델이 수행하는 작업은 Curie 또는 Davinci와 같은 더 강력한 모델이 수행할 수 있습니다.

For tasks involving classification, where you’re trying to find what label best fits a selection of text, you can often get great performance out of the different models with Semantic Search. Semantic Search uses an models to provide scores for different blocks of text for how closely they relate to a query. By focusing the scope of the model to evaluating the relationship of the query to the different text blocks, in many cases faster models can outperform themselves compared to the task being presented to them as a generative one.

선택한 텍스트에 가장 적합한 레이블을 찾으려는 분류 관련 작업의 경우 시맨틱 검색을 통해 다양한 모델에서 뛰어난 성능을 얻을 수 있습니다. 시맨틱 검색은 모델을 사용하여 쿼리와 얼마나 밀접하게 관련되어 있는지에 대한 다양한 텍스트 블록에 대한 점수를 제공합니다. 다른 텍스트 블록에 대한 쿼리의 관계를 평가하는 데 모델의 범위를 집중함으로써 많은 경우 더 빠른 모델이 생성 작업으로 제시되는 작업에 비해 더 빠른 모델을 능가할 수 있습니다.