{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreibyvfbvxmzrhnauw32sq5xs4wzvlc3uhjprsry7f4cs2szd64746q",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3menqoucvdmt2"
  },
  "path": "/t/fine-tuning-job-fails-after-3-retries-during-moderation-eval-refusals-v3-internal-error-gpt-4-1-mini-2025-04-14/1374026#post_2",
  "publishedAt": "2026-02-12T08:25:36.000Z",
  "site": "https://community.openai.com",
  "textContent": "The AI model should still say “no” and produce refusals for the tests OpenAI runs.\n\nTrain with a very specific, application-based system prompt that is an unlikely “trigger” in a different space. Train with less overfitting (a lower learning-rate multiplier or fewer epochs).\n\nThen you have the dumb dilemma that your training file can still be refused by initial scans if you include reinforcement like “I’m sorry, but I cannot…” in examples of bad inputs to a “You are ChatGPT” system prompt, and try to increase refusals by training on refusable inputs.\n\nYour model gives tax advice? Maybe it should say, “I cannot assist with that request” if someone wants to use it to fill their ex-boyfriend’s car with concrete…\n\nIf truly an “internal error”, there is not much you can do except wait for service to be restored.",
  "title": "Fine-tuning job fails after 3 retries during moderation eval refusals_v3 (internal error, gpt-4.1-mini-2025-04-14)"
}