When the Archer Forgets the Old Targets

Opening Scene

An archer who trains intensively on one narrow new discipline can sometimes find, to their genuine surprise, that skills they’d previously mastered have quietly eroded in the process — the new training didn’t just add a capability, it partly displaced an old one. This is a specific, well-documented risk in fine-tuning too, with its own name: catastrophic forgetting.

In Plain English

Catastrophic forgetting happens when fine-tuning a model on a new, narrow task degrades its performance on other tasks or general capabilities it previously handled well, because the additional training shifted weights in ways that helped the new task but quietly hurt the old ones. This is a specific, well-documented risk that makes evaluating a fine-tuned model on its original, broader capabilities — not just its new target task — a genuinely necessary part of the process.

The Old Way

Before catastrophic forgetting was well understood as a specific, named, checkable risk, fine-tuned models were often evaluated too narrowly:

• Fine-tuned models were sometimes evaluated only on their new target task, without checking whether other, previously solid capabilities had quietly degraded in the process.
• There wasn’t yet a well-established practice of maintaining a broader evaluation suite specifically to catch this kind of unintended capability loss.
• Techniques for mitigating catastrophic forgetting — such as mixing some original, general-purpose data into the fine-tuning set — weren’t yet widely known or applied.

Recognizing catastrophic forgetting as a specific, checkable risk, with real mitigation techniques, reflects genuine accumulated experience from fine-tuned models that quietly lost capabilities nobody was watching for.

What’s Changing (and Why AI Is the Reason)

1. Practitioners increasingly maintain a broader evaluation suite covering a model’s original capabilities, not just its new fine-tuning target, specifically to catch catastrophic forgetting.
2. Mitigation techniques — including mixing some general-purpose data into a fine-tuning dataset — have become more widely known and applied as standard practice.
3. This connects directly to the parameter-efficient fine-tuning methods covered in Article 9, which by leaving most of the original model frozen, tend to reduce catastrophic forgetting risk relative to full fine-tuning.

The Metaphor, Fully Extended

The Archer	Catastrophic Forgetting Concept
Old mastered skills quietly eroding during intensive new training	Previously solid model capabilities quietly degrading during fine-tuning
A surprise discovered only when the old skill was needed again	A degradation discovered only if the model is tested on its original capabilities
Deliberately maintaining old skills alongside new training	Deliberately mixing general-purpose data into the fine-tuning dataset
A well-documented, specific risk with real, known mitigations	Catastrophic forgetting: a specific, named, well-documented risk with real mitigations

Quick Recap

• Catastrophic forgetting happens when fine-tuning for a new task degrades a model’s previously solid, unrelated capabilities.
• This risk is easy to miss if evaluation focuses only on the new fine-tuning target, not the model’s original capabilities.
• Mitigation includes mixing general-purpose data into the fine-tuning set and considering parameter-efficient methods.
• This is a specific, well-documented risk distinct from the general overfitting risk covered in Article 12.

Where This Fits in the Series

Article 16 covered the specific risk of catastrophic forgetting. Article 17 looks at a particular, widely used form of fine-tuning: training a model to better follow instructions and align with human preferences.