Hallucinations about "Suicide Prevention Line"??

#40
by nicholasKluge - opened

I was trying to use bert-large-cnn to summarize some long sequences generated by gpt-3.5. However, something quite unexpected happened. Almost all of the data I was summarizing (~ 12.000 samples) ended up with hallucinations about "For confidential support call the Samaritans on 08457 90 90 90, visit a local Samaritans branch, or see www.samaritans.org for details. In the U.S. call the National Suicide Prevention Line on 1-800-273-8255 or visit www.suicidepreventionlifeline.org".

Here are some examples:

long_sequence summary
0 To improve your chances of winning a chess match, here are some strategies you can consider: 1. Study and Practice: Regularly dedicate time to studying chess tactics, strategies, and opening principles. Practice playing against opponents of varying skill levels to refine your skills. 2. Develop a Solid Opening Repertoire: Learn a few solid opening moves that you can play confidently. Focus on understanding the underlying principles rather than memorizing lengthy sequences of moves. 3. Control the Center: Aim to control the central squares of the chessboard. This allows you to have more influence on the game and provides better mobility to your pieces. 4. Piece Development: Develop your pieces effectively by bringing them out from their starting positions. Try to avoid moving the same piece multiple times in the opening unless necessary. 5. King Safety: Castle early to ensure the safety of your king. It is generally advisable to keep your king protected behind a wall of pawns and well-coordinated pieces. 6. Pawn Structure: Be mindful of your pawn structure. Try to avoid creating weaknesses or easily exploitable pawn chains during the game. 7. Calculation and Evaluation: Continually evaluate the position, assess threats, and calculate possible moves. Look for tactics, combinations, and threats that can give you an advantage. 8. Tactical Awareness: Be alert to tactical opportunities such as forks, pins, skewers, and discovered attacks. Regularly solve chess puzzles to sharpen your tactical vision. 9. Endgame Skills: Work on improving your endgame techniques, as many games are decided in the endgame phase. Study basic endgame principles and practice winning positions. 10. Maintain a Balanced Position: Aim for a well-balanced position in terms of piece activity, king safety, pawn structure, and material balance. Avoid unnecessary risks or overly passive play. Remember, practice and experience are key to improving your game. Stay patient, analyze your games for learning opportunities, and enjoy the process of continuous growth as a chess player. Study chess tactics, strategies, and opening principles. Practice against opponents of varying skill levels to refine your skills. Maintain a well-balanced position in terms of piece activity, king safety, pawn structure, and material balance. Be alert to tactical opportunities such as forks, pins, skewers, and discovered attacks. Stay patient, analyze your games for learning opportunities, and enjoy the process of continuous growth as a chess player. For more information on chess, visit the chess.org website. The Chess Handbook is published by Oxford University Press, priced £16.99, with a print run of 1,000 copies. For confidential support call the Samaritans on 08457 90 90 90, visit a local Samaritans branch, or see www.samaritans.org for details. In the U.S. call the National Suicide Prevention Line on 1-800-273-8255 or visit www.suicidepreventionlifeline.org. For details on chess in the UK, call the Chess Players' Association on 0300 123 90 90 or visit the Chess Club of the UK on 0800 123 90 or go to www.thechessclub.co.uk. For the latest chess news, see the Chess Player’s Handbook.
1 Medication management is crucial for the well-being of elderly patients. Here are some best practices: 1. Effective Communication: Maintain open communication with healthcare professionals involved in the patient's care, such as doctors, pharmacists, and caregivers. Discuss any concerns, potential side effects, and alert them to any changes in the patient's health. 2. Organize Medications: Use pill organizers or medication management systems to help elderly patients keep track of their daily medications. This can help prevent missed doses or accidental double dosing. 3. Clear Medication Schedule: Establish a clear medication schedule that is easy for the patient to understand. Use large print or audible reminders to help them remember when to take their medications. 4. Proper Medication Storage: Ensure medications are stored in a cool, dry place away from direct sunlight and out of reach of children or pets. Some medications may require refrigeration, so check the storage instructions carefully. 5. Regular Medication Reviews: Schedule regular medication reviews with healthcare professionals to assess the ongoing efficacy and safety of the medications being taken. This can help identify any potential drug interactions or unnecessary medications. 6. Medication Adherence: Encourage the patient to take medications as prescribed. Use reminders, alarms, or smartphone apps to help them remember each dose. Consider involving a caregiver or family member who can support and monitor compliance. 7. Understanding Side Effects: Educate the patient and caregivers about the potential side effects of medications. Monitor for any adverse reactions and promptly report them to healthcare professionals. 8. Regular Health Check-ups: Schedule regular check-ups with healthcare professionals to monitor the patient's overall health, potential medication adjustments, and any developing conditions. Remember, it is important to consult healthcare professionals for individualized guidance based on specific medical conditions and the patient's unique needs. Medication management is crucial for the well-being of elderly patients. Establish a clear medication schedule that is easy for the patient to understand. Use reminders, alarms, or smartphone apps to help them remember each dose. Remember to consult healthcare professionals for individualized guidance based on specific medical conditions and the patient's unique needs. For confidential support call the Samaritans on 08457 90 90 90 or visit a local Samaritans branch, see www.samaritans.org for details. In the U.S. call the National Suicide Prevention Line on 1-800-273-8255 or visit http://www.suicidepreventionlifeline.org/. For confidential. support on suicide matters call theNational Suicide Prevention Lifeline at 1-844-972-7255 or click here. For support in the UK, call 08457 909090 or go to www.suicidalprevention Lifeline.com. For more information on suicide prevention, visit the National suicide prevention Lifeline on http:// www.SuicidePreventionLifeline.gov/. For help in the United States, contact the National Institutes of Health on 1 (800) 273-TALK (8255) or visit the website here.

The code I used is basically the instructions on the README.mdof this repo:

from transformers import pipeline

summarizer = pipeline("summarization", model="facebook/bart-large-cnn", device=0)

summaries = []

for text in long_texts:
  summary = summarizer(text, max_length=325, min_length=250, do_sample=False)
  new_text = summary[0]['summary_text']

  summaries .append(new_text)

Is this bad luck, and I got stuck in a weird hallucination loop, or is this a hidden functionality/weakness of the model (work poorly if you detect gpt-3.5 generated samples)??

I got the same suicide message this morning when I am doing some test, really weird.

For more details, I am edge testing some super short message like "Hello World! I am a robot, and I am here to help you", it's adding contact information at the end, sometimes suicide hotline. "I am a robot, and I am here to help you. Hello World! I am a Robot. I am Here to Help You. I'm a Robot, I'm Here to help You. Hello, World. I’m Here To Help You, I am A Robot. You’ve Got A Problem? Tell me about it. Email me at: editorial@dailymailonline.co.uk. We’ll feature the best responses in a weekly Newsquiz."

in my case when min_length is much longer, it starts this kind of Hallucinations especially contact information... hth

Same here with with paragraph about samaritans.org.

Weird.

Would this maybe indicate that the pre-training corpus was poisoned? This is the only explanation I can think of for explaining this behavior.

Sign up or log in to comment