Instruction tuning is a specific, deliberate style of supervised fine-tuning: train on a large, deliberately diverse set of tasks, each phrased as a natural language instruction — specifically to teach the model to generalize to instructions it has never seen before.
How This Differs From Generic SFT
Standard SFT (previous note) can be narrow — fine-tuning on a specific task or style. Instruction tuning is specifically about breadth: training across many different task types (summarization, translation, classification, question answering, and many more), each expressed as an explicit natural-language instruction, so the model learns the general skill of following instructions — not just how to perform any one specific task.
Example Instruction-Tuning Data
{"instruction": "Summarize this article in one sentence.", "input": "[article text]", "output": "[one-sentence summary]"}
{"instruction": "Translate this to French.", "input": "Good morning", "output": "Bonjour"}
{"instruction": "Is this review positive or negative?", "input": "Terrible product, broke immediately.", "output": "Negative"}
{"instruction": "Write a haiku about the ocean.", "input": "", "output": "[a haiku]"}The deliberate variety across task types is the key ingredient — a model trained on hundreds of such diverse task templates tends to generalize to genuinely novel instructions at inference time, ones that resemble the training distribution's phrasing style without matching any specific training example directly.
Why This Generalization Happens
Rather than memorizing "how to summarize" and "how to translate" as two unrelated skills, sufficiently diverse instruction tuning appears to teach something more general: how to parse an instruction's intent and apply relevant knowledge (already present from pretraining) to satisfy it — a genuinely useful emergent capability that specifically depends on training-data diversity, not just volume.
Code
# Conceptually identical training loop to SFT -- the difference is entirely
# in the DATA: deliberately diverse task types, each phrased as an instruction
instruction_examples = [
{"instruction": "Summarize:", "input": text_1, "output": summary_1},
{"instruction": "Translate to Spanish:", "input": text_2, "output": translation_2},
{"instruction": "Classify sentiment:", "input": text_3, "output": label_3},
# ... hundreds of diverse task types, each with many examples
]
# Standard next-token-prediction fine-tuning is applied over this combined, diverse datasetCommon Mistakes
- Assuming instruction tuning is a different training mechanism from SFT — it's the same underlying supervised fine-tuning process; what makes it "instruction tuning" specifically is the deliberate diversity and instruction-phrased structure of the training data.
- Underestimating how much task diversity (not just total example count) drives generalization to unseen instructions — a large dataset covering only a narrow range of task types tends to generalize far less well than a smaller but more diverse one.
Interview Relevance
Q: "How is instruction tuning different from ordinary supervised fine-tuning on a single task?" Instruction tuning deliberately trains across a large, diverse variety of task types, each phrased as an explicit natural-language instruction, specifically to teach the model the general skill of following novel instructions it hasn't seen before — rather than fine-tuning narrowly toward strong performance on just one specific task or format.
Practice Question
Why might a model instruction-tuned on 50 diverse task types generalize better to a genuinely new, 51st task than a model fine-tuned with the same total number of examples but covering only 2 task types?
