Your fine-tuned model produces factually wrong outputs due to training data quality issues.
LLM & AILevel: IntermediateTroubleshooting Scenarios
Your fine-tuned model produces factually wrong outputs due to training data quality issues.
Hints
- Which parameters change and which stay frozen?
- Separate memory for weights, gradients, optimizer states and activations.
Answer
- Audit training data — sample 100 examples; rate factual accuracy manually; compute error rate
- Identify systematic errors — are errors in specific topics, formats, or data sources?
- Data cleaning — remove or correct factually wrong examples; add factual verification step to data pipeline
- RAG hybrid — don't rely on fine-tuned weights for factual recall; add RAG to ground outputs in verified sources
- Confidence calibration — add training examples where the correct answer is "I don't know" for uncertain domains
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0