A student fine-tuned a model hard on customer-support replies and it got noticeably worse at basic reasoning questions it had handled fine before. That's catastrophic forgetting — narrow, aggressive fine-tuning overwrites general capability the base model had.
Why it happens
Too many epochs on too narrow a dataset pushes the model's weights hard in one direction. LoRA reduces this risk versus full fine-tuning because it touches far fewer parameters, but it isn't immune, especially at high rank and learning rate.
How to prevent it
- Evaluate on a general-capability set before and after fine-tuning, not just on the target task.
- Fewer epochs and a lower learning rate first; add more only if the target task genuinely needs it.
- Mix a small slice of general instruction data into the fine-tuning set alongside the task-specific data.
See reading a loss curve during fine-tuning for how forgetting often shows up as a training loss that looks fine while eval quality quietly drops elsewhere.
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU Kanpur) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. Email pranjulrathour41@gmail.com.
