What is the difference between a base model, an instruction-tuned model and a chat model?

// Base model vs chat model, in miniature. The "model" is a back-off n-gram
// (it predicts the next token from the longest context it has seen), so the
// same code shows how pretraining data and fine-tuning data change behaviour.
const MAX_CTX = 6;

function train(texts) {
  const table = {};
  for (const text of texts) {
    const t = text.split(" ");
    for (let i = 1; i < t.length; i++)
      for (let k = 1; k <= Math.min(MAX_CTX, i); k++) {
        const ctx = t.slice(i - k, i).join(" ");
        table[ctx] = table[ctx] || {};
        table[ctx][t[i]] = (table[ctx][t[i]] || 0) + 1;
      }
  }
  return table;
}

function generate(table, prompt, maxNew, stop) {
  const out = prompt.split(" ");
  for (let n = 0; n < maxNew; n++) {
    let counts = null;
    for (let k = Math.min(MAX_CTX, out.length); k >= 1 && !counts; k--) counts = table[out.slice(-k).join(" ")] || null;
    if (!counts) return { text: out.slice(prompt.split(" ").length).join(" ") || "(nothing)", why: "no known context" };
    const next = Object.keys(counts).sort().reduce((a, b) => (counts[b] > counts[a] ? b : a));
    if (next === stop) return { text: out.slice(prompt.split(" ").length).join(" "), why: "stop token " + stop };
    out.push(next);
  }
  return { text: out.slice(prompt.split(" ").length).join(" "), why: "hit max " + maxNew + " tokens" };
}

// Stage 1, pretraining: raw web-style text. Nobody is "answering" anything.
const webText = [
  "what is the capital of france ? what is the capital of spain ? what is the capital of italy ? " +
  "quiz books list questions like these . paris is the capital of france . " +
  "madrid is the capital of spain . rome is the capital of italy .",
];

// Stage 2, instruction tuning: conversations rendered with a chat template.
// Special tokens mark who is speaking and where a turn ends.
const render = (messages, addGenerationPrompt) =>
  messages.map((m) => "<|im_start|> " + m.role + " " + m.content + " <|im_end|>").join(" ") +
  (addGenerationPrompt ? " <|im_start|> assistant" : "");
const sftPairs = [
  ["what is the capital of france ?", "paris is the capital of france ."],
  ["what is the capital of spain ?", "madrid is the capital of spain ."],
];
const sftTexts = sftPairs.map(([q, a]) => render([{ role: "user", content: q }, { role: "assistant", content: a }], false));
const EPOCHS = 3; // fine-tuning repeats the small dataset, shifting the counts

const base = train(webText);
const chat = train([...webText, ...Array(EPOCHS).fill(sftTexts).flat()]);
console.log("one training conversation as the model sees it:");
console.log("  " + sftTexts[1]);

const question = "what is the capital of spain ?";
let r = generate(base, question, 14, "<|im_end|>");
console.log("BASE, raw prompt   -> " + r.text + "   [" + r.why + "]");

const prompt = render([{ role: "user", content: question }], true);
console.log("chat prompt string -> " + prompt);
r = generate(base, prompt, 14, "<|im_end|>");
console.log("BASE, chat prompt  -> " + r.text + "   [" + r.why + "]");
r = generate(chat, prompt, 14, "<|im_end|>");
console.log("CHAT, chat prompt  -> " + r.text + "   [" + r.why + "]");

const ctx = "is the capital of spain .";
const show = (t) => Object.keys(t[ctx]).sort().map((w) => w + ":" + t[ctx][w]).join(" ");
console.log("after '" + ctx + "'  base counts: " + show(base) + " | chat counts: " + show(chat));