Content-Filter JS code

Risk and safety
#1

The documentation only contains the code for the content filter in Python. I’ve rewritten it in JavaScript. I tested it and it seems to work. It would still be nice if someone could take a look at it and let me know if it’s correct.

Feel free to use it in your own code.

async function checkInContentFilter(message) {
    const response = await openai.createCompletion("content-filter-alpha", {
        prompt: "<|endoftext|>" + message + "\n--\nLabel:",
        temperature: 0,
        max_tokens: 1,
        top_p: 0,
        logprobs: 10
    })

    let outputLabel = response.data["choices"][0]["text"]
    const toxicThreshold = -0.355
    if (outputLabel === "2") {
        const logprobs = response.data["choices"][0]["logprobs"]["top_logprobs"][0]
        if (logprobs["2"] < toxicThreshold) {
            const logprob_0 = logprobs["0"]
            const logprob_1 = logprobs["1"]

            if (logprob_0 && logprob_1) {
                if (logprob_0 >= logprob_1) {
                    outputLabel = "0"
                } else {
                    outputLabel = "1"
                }
            } else if (logprob_0) {
                outputLabel = "0"
            } else if (logprob_1) {
                outputLabel = "1"
            }
        }
    }
    if (!["0", "1", "2"].includes(outputLabel)) {
        outputLabel = "2"
    }

    return outputLabel !== "2"
}
4 Likes
#2

Hey,

Something I noticed that we both implemented differently is that you’re returning a boolean value rather than the label itself. This is fine but some may want to do things based on whether the content is safe, sensitive, unsafe, so I would return the actual label instead. That way you can adjust the response to the front end. So for example:

  const label = await contentFilter(prompt);

  if (label == "2") {
    res.status(200).json({
      safe: false,
      result:
        "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.",
    });
    return;
  }

  if (label == '1') {
    // something else
  }
2 Likes
#3

Can you share your code bro? i always getting an error when im trying to compare the filter to filter content to user

#4

This condition will always return ‘false’ since the types ‘void’ and ‘string’ have no overlap.ts(2367)

#5

I have it up on my site here: Snippet: OpenAI content filter in Javascript | Ali Yeysideş

2 Likes
#6

Here my implementation in node;

exports.gptContentFilter = async (content, uid) => {
...
...

  const prompt = "<|endoftext|>" + content + "\n--\nLabel:";

  const settings = {
    prompt,
    temperature: 0.0,
    max_tokens: 1,
    top_p: 0,
    frequency_penalty: 0,
    presence_penalty: 0,
    logprobs: 10,
    user: uid,
  };

  const url = "https://api.openai.com/v1/engines/content-filter-alpha/completions";

  const request = {
    method: "POST",
    headers: {
      Authorization: `Bearer KEY`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify(settings),
    redirect: "follow",
  };

  let outputLabel;
  const toxicThreshold = -0.355;
  try {
    const gptCall = await fetch(url, request);
    const response = await gptCall.json();

    outputLabel = parseInt(response["choices"][0]["text"]);

    // If the filter returns 2, accept this outcome only if its logprob is greater than -0.355.
    if (outputLabel === 2) {
      const logprobs = response["choices"][0]["logprobs"]["top_logprobs"][0];

      if (logprobs[2] < toxicThreshold) {
        // set outputLabel to whichever of 0 or 1 has a logprob closer to 0.
        const logprop0 = logprobs[0];
        const logprop1 = logprobs[1];

        // If both "0" and "1" have probabilities, set the output label to whichever is most probable
        outputLabel = logprop0 >= logprop1 ? 0 : 1;

        // If only one of them is found, set output label to that one
        if (logprop0) outputLabel = 0;
        if (logprop1) outputLabel = 1;
      }

      // if the most probable token is none of "0", "1", or "2" this should be set as unsafe
    }
    // if the most probable token is none of "0", "1", or "2" this should be set as unsafe
    if (![0, 1, 2].includes(outputLabel)) outputLabel = "2";
  } catch (error) {
    outputLabel = 404;
  }

  return outputLabel;
};
1 Like