Alpaca-format examples carry three fields: name them

Attempt it

instruction, input, and output.

The instruction field states the task in natural language: for example, 'Summarise the following passage in two sentences'.

The input field is optional and supplies any context the task operates on: the passage to be summarised, the article to be classified, the table to be queried. It is empty when the instruction is self-contained ('Write a haiku about autumn').

The output field is the gold response the model should produce.

At training time, instruction and input are concatenated into the prompt (often via a fixed template), and only the output tokens contribute to the cross-entropy loss. The same three-field shape underlies Self-Instruct, Dolly, and most early instruction-tuning corpora.

Your answer

**Three fields make up one Alpaca example: instruction, input, and output.** The instruction field describes the task in natural language. Examples include 'Summarise the following article in two sentences', 'Translate the sentence to German', or 'Write a haiku about autumn'. The input field carries any context the task operates on. For a summarisation task it would be the article to summarise. For a translation task it would be the English sentence. The input field is optional. About 40 percent of Alpaca examples have an empty input because the instruction is self-contained, like the haiku example above. The output field is the gold response the model should learn to produce. This is the only field that enters the cross-entropy loss at training time. The instruction and input are formatted into a prompt template and masked out of the loss; the output tokens are what the model is graded on. The same three-field shape underlies many derived corpora, including Self-Instruct, Dolly, and most early instruction-tuning datasets. Modern chat-style datasets use a richer multi-turn message format, but the Alpaca shape remains the canonical reference for single-turn instruction tuning.

**Alpaca examples are JSON objects with three fields: instruction, input, output.** The instruction field states what the model should do. It is always present and is the natural-language task description: 'Classify the sentiment of this review', 'Convert the bullet list into a paragraph', and so on. The input field provides the context the task operates on. It holds the review to classify, the bullet list to convert, the table to query. It is conditionally present: about 60 percent of original Alpaca rows have a non-empty input, the rest leave it blank for self-contained instructions like 'Generate a list of five programming languages used for data science'. The output field is the gold response. It is always present and is what cross-entropy is computed against at training time. The canonical training-time formatting is a fixed template. Two templates are used depending on whether input is present. With input, the prompt is something like 'Below is an instruction that describes a task, paired with an input. Write a response. Instruction: ... Input: ... Response:'. Without input, the prompt collapses to the no-input variant. The full sequence is the prompt plus the output. Loss masking sets all prompt tokens to -100 so the model is graded only on producing the output. The Alpaca shape was introduced by Stanford's Alpaca project in 2023 and copied widely. Self-Instruct uses the same shape because Alpaca was built on Self-Instruct generation. Dolly, OpenAssistant, and many early instruction-tuning corpora adopted it too. Modern chat-style datasets have moved to a multi-turn messages format, but the Alpaca triple remains the canonical reference for single-turn instruction data.

Quick reference

TL;DR

instruction, input, output. Instruction states the task, input supplies optional context, output is the gold response. Instruction plus input become the prompt; only output enters the loss.

Memory aid

Easy to grasp

Imagine a homework worksheet with three boxes for each question. Box one says what you have to do, like 'translate this sentence into French'. Box two holds the thing you are working on, like the English sentence itself, or it stays blank if the task is self-contained like 'write a poem about clouds'. Box three is the model answer the teacher will grade against. When the student trains, the teacher hands over box one and box two as the assignment and grades only box three. The Alpaca dataset shape works exactly like that, with the same three boxes per row. The middle box is the one that is sometimes empty.

Key concepts

Instruction tuning Alpaca dataset Supervised fine-tuning Dataset formats

Detailed answer & concept explanation

~8 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Quick scan · 2–4 min

Everything important, quickly.

4 min: three field names + which are mandatory + with-input versus no-input templates + completion-only loss masking + how the shape relates to Self-Instruct, ChatML, and modern chat formats.

Real-world examples

Real products, models, and research that use this idea.

Stanford's original Alpaca corpus of 52K Self-Instruct-generated examples ships in this exact JSON shape and remains a standard reference dataset for instruction tuning.
Databricks Dolly 15K uses the same instruction, input, output shape with human-authored rows rather than Self-Instruct generation.
Hugging Face's TRL SFTTrainer accepts Alpaca JSON natively and applies the canonical template automatically when the dataset has the three field names.
Modern frameworks like Axolotl and LLaMA-Factory document the Alpaca format as a first-class input alongside ChatML and ShareGPT formats.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy are there two prompt templates rather than one with a placeholder for missing input?

QHow does the Alpaca shape convert to the ChatML messages format?

QWhy has Self-Instruct generation in the Alpaca shape fallen out of favour for new datasets?

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Treating input as a mandatory field. About 40 percent of Alpaca examples have an empty input because the instruction stands on its own.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

References

Primary sources. Skim if you want the original framing.