The Coach Reviewing Form

Opening Scene

Before an archer ever competes, a coach spends real time reviewing their raw form and correcting it toward good, generally applicable technique — not for any one specific target, but so the archer behaves well and predictably across the board. This is exactly the role instruction tuning and reinforcement learning from human feedback play for large language models: a specific, foundational kind of fine-tuning that happens before a model is ever handed to a user for a specific task.

In Plain English

A raw, pretrained language model is genuinely capable but not naturally well-behaved as an assistant — it doesn’t reliably follow instructions or align well with what people actually want by default. Instruction tuning and reinforcement learning from human feedback (RLHF) are themselves specific forms of fine-tuning applied to shape a raw pretrained model into the genuinely helpful, instruction-following assistant most people interact with. This is why the “base model” most fine-tuning and prompting work builds on top of has already been fine-tuned once, for this general purpose, before task-specific work even begins.

The Old Way

Before instruction tuning and RLHF were widely applied, the gap between raw pretrained capability and genuinely usable assistant behavior was much more visible:

• Early raw pretrained language models were significantly harder to use directly, often continuing text in unhelpful or unexpected ways rather than following instructions reliably.
• There wasn’t yet a well-established practice of fine-tuning specifically for instruction-following and alignment with human preferences, as a distinct step from task-specific fine-tuning.
• The distinction between a model’s raw pretrained capability and its instruction-following, assistant-like behavior wasn’t yet a widely understood, separate layer.

Recognizing instruction tuning and RLHF as this foundational fine-tuning layer clarifies why today’s models are already so much more usable, straight out of the box, than early raw pretrained models were.

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

1. Instruction tuning and RLHF have become a standard, foundational step applied to nearly all widely used language models, before any task-specific fine-tuning or prompting happens on top.
2. This connects directly to this content library’s dedicated series on AI governance and responsible AI, since RLHF specifically shapes a model’s alignment with human preferences and safety considerations.
3. Understanding this layer clarifies that “the model” most practitioners work with has already been fine-tuned once for general usability, which is part of why task-specific fine-tuning today needs comparatively less additional data.

The Metaphor, Fully Extended

The Archer	Instruction Tuning / RLHF Concept
A coach reviewing raw form before any specific competition	Instruction tuning and RLHF shaping a raw model before task-specific use
General, predictable good technique, not tuned for one target	General instruction-following and alignment, not tuned for one task
A foundational layer every subsequent training builds on	A foundational fine-tuning layer every subsequent task-specific work builds on
An archer who’s already well-coached before specialized training begins	A base model that’s already instruction-tuned before task-specific fine-tuning begins

Quick Recap

• Instruction tuning and RLHF are themselves forms of fine-tuning, applied to shape a raw pretrained model into a genuinely usable assistant.
• This is a foundational layer that happens before any task-specific fine-tuning or prompting takes place.
• This connects directly to alignment and responsible AI considerations covered in this content library’s dedicated governance series.
• Understanding this layer clarifies why today’s base models need comparatively less task-specific data to fine-tune well.

Where This Fits in the Series

Article 17 covered the foundational fine-tuning layer beneath every model. Article 18 pulls every factor covered so far into one practical decision framework for choosing between prompting and fine-tuning on a real project.