0.3% In, 36% Out: Your Fine-Tuned Model Is Copying Your Prompt Examples
I run a fine-tuned Llama 3.3 70B on Amazon Bedrock. It generates short first-person narrative posts: a setup, a few lines of story, and a closing line that lands the point. Last week I noticed the closing lines had collapsed into a single template. Not similar in spirit, but literally the same grammatical shape, over and over: and that's how [someone] [learns/teaches] [something]. Roughly a third…
A fine-tuned Llama 3.3 70B model on Amazon Bedrock started producing short narrative posts with identical closing lines. The model's output was heavily influenced by the training data, with 36% of generated posts ending with the same template. The obvious solution was to retrain the model with better data, but before spending resources, the reporter conducted a thorough analysis.
By counting the occurrence of the problematic template in the training data and the generated posts, the reporter discovered that the template appeared only five times in 1,610 training examples (0.3%), but 36% of the generated posts (9 out of 25) followed the same structure. The reporter realized that the model had essentially copied the example from the prompt, rather than learning from the training data.
This realization led the reporter to fix the issue by diversifying the closing lines in the prompt, rather than retraining the model. The reporter recommends persisting generated examples and treating prompt examples as training data to prevent similar issues in the future.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
