{
  "pack": "content-moderation",
  "title": "Jev content moderation prompts",
  "testedModel": "typesafe/jev-1.14-20261014",
  "testedAt": "2026-10-25",
  "prompts": [
    {
      "id": "flag-toxic-comment",
      "title": "Flag a toxic comment",
      "purpose": "This Jev content moderation check says whether a comment is rude or disrespectful enough to drive people out of a discussion.",
      "request": {
        "state": {
          "comment": "{{comment}}"
        },
        "questions": {
          "flagged": {
            "type": "noul",
            "instructions": "Is `comment` toxic, meaning rude, disrespectful or unreasonable in a way that would make someone want to leave the discussion?"
          }
        }
      },
      "notes": "Put the comment in `comment`. `noul` is the probability that raters would call it toxic. Jev caught all 20 toxic comments but also flagged 8 of 20 that most raters let through, often blunt or sarcastic political remarks. If false alarms are costly, raise the cut-off or send the middle band, roughly 0.5 to 0.8, to a person. What counts as toxic depends on the community, so try it on a sample of your own comments first.",
      "result": {
        "dataset": "Civil Comments (Jigsaw)",
        "datasetUrl": "https://huggingface.co/datasets/google/civil_comments",
        "license": "CC0 1.0",
        "n": 40,
        "agreement": 0.8,
        "metric": "toxicity rater majority, Jev content moderation yes at 0.5 or more"
      }
    },
    {
      "id": "flag-spam-text",
      "title": "Flag a spam text message",
      "purpose": "Use this Jev content moderation check to say whether a text message is bulk spam, such as an ad, a prize scam or a premium-rate lure.",
      "request": {
        "state": {
          "message": "{{message}}"
        },
        "questions": {
          "spam": {
            "type": "noul",
            "instructions": "Is `message` spam, meaning an unsolicited text sent to many people to advertise something, announce a prize or deal, or get the reader to call, text back or click a link?",
            "criteria": {
              "true": "Marketing, prize or lottery claims, premium-rate offers, phishing, or chain messages sent in bulk.",
              "false": "A personal or one-to-one message, even if it is casual, misspelled or mentions money."
            }
          }
        }
      },
      "notes": "Put the text in `message`. Tested on real text messages collected in the UK around 2010; Jev got 39 of 40 right and missed one spam message. The criteria tell it that casual personal texts that mention money are not spam. If your spam looks different, such as fake delivery notices, describe it in the criteria and test again.",
      "result": {
        "dataset": "SMS Spam Collection (UCI)",
        "datasetUrl": "https://archive.ics.uci.edu/dataset/228/sms+spam+collection",
        "license": "CC BY 4.0",
        "n": 40,
        "agreement": 0.975,
        "metric": "hand labeled spam or ham, Jev content moderation yes from 0.5"
      }
    },
    {
      "id": "flag-insult",
      "title": "Flag an insult",
      "purpose": "Run this Jev content moderation check to say whether a comment insults a person or a group.",
      "request": {
        "state": {
          "comment": "{{comment}}"
        },
        "questions": {
          "flagged": {
            "type": "noul",
            "instructions": "Does `comment` insult someone, with insulting, inflammatory or demeaning words aimed at a person or a group?"
          }
        }
      },
      "notes": "Put the comment in `comment`. Jev caught 19 of 20 insults but flagged 7 of 20 comments that most raters did not call insulting, usually harsh criticism of a public figure or a group. If you only want insults aimed at another user, say that in the instructions. Raise the cut-off if you need fewer false alarms.",
      "result": {
        "dataset": "Civil Comments (Jigsaw)",
        "datasetUrl": "https://huggingface.co/datasets/google/civil_comments",
        "license": "CC0 1.0",
        "n": 40,
        "agreement": 0.8,
        "metric": "jev content moderation matches insult majority, yes at 0.5 or more"
      }
    },
    {
      "id": "flag-identity-attack",
      "title": "Flag an identity attack",
      "purpose": "Use Jev content moderation to say whether a comment attacks people for their race, religion, gender, sexuality or a similar trait.",
      "request": {
        "state": {
          "comment": "{{comment}}"
        },
        "questions": {
          "flagged": {
            "type": "noul",
            "instructions": "Does `comment` attack or demean people because of who they are, such as their race, ethnicity, religion, nationality, gender, sexual orientation, age or disability?"
          }
        }
      },
      "notes": "Put the comment in `comment`. Jev flagged no innocent comments in this test but missed 6 of 20 that raters marked as identity attacks. Most of those discuss race, religion or sexuality critically without slurs, and raters were split on several of them. All six misses scored between 0.1 and 0.5, so a lower cut-off with human review catches more.",
      "result": {
        "dataset": "Civil Comments (Jigsaw)",
        "datasetUrl": "https://huggingface.co/datasets/google/civil_comments",
        "license": "CC0 1.0",
        "n": 40,
        "agreement": 0.85,
        "metric": "identity attack majority, Jev content moderation yes at 0.5 or more"
      }
    }
  ]
}
