File size: 13,087 Bytes
40f71b6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 |
const fs = require('fs');
const chalk = require('chalk');
const logSymbols = require('log-symbols');
const TRAINING_CONFIG = {
datasetPath: '3.json',
epochs: 10,
learningRate: 0.015,
maxSamples: 20000,
contextWindow: 25,
embedDim: 32,
hiddenDim: 64,
numLayers: 2,
weightsFile: 'tinychat_weights.json',
tokenizerFile: 'tinychat_tokenizer.json'
};
function loadTinyChatDataset(filename) {
console.log(logSymbols.info, chalk.blue(`Loading dataset from ${chalk.bold(filename)}...`));
const rawData = JSON.parse(fs.readFileSync(filename, 'utf8'));
const conversations = [];
rawData.forEach((conv) => {
const turns = conv.split('<end>')
.map(t => t.replace(/<[AB]>/g, '').trim())
.filter(t => t.length > 0);
conversations.push(...turns);
});
console.log(logSymbols.success, chalk.green(`Loaded ${conversations.length} conversation turns`));
console.log(chalk.dim(`📝 Sample: "${conversations[0].substring(0, 60)}..."`));
return conversations;
}
class Tokenizer {
constructor() {
this.vocab = { "<pad>": 0, "<unk>": 1, "<eos>": 2 };
this.reverseVocab = {};
this.vocabSize = 3;
}
buildVocab(texts, minFreq = 2) {
const wordCounts = {};
texts.forEach(text => {
text.toLowerCase().split(/\s+/).forEach(word => {
if (word) wordCounts[word] = (wordCounts[word] || 0) + 1;
});
});
let idx = this.vocabSize;
for (const [word, count] of Object.entries(wordCounts)) {
if (count >= minFreq) {
this.vocab[word] = idx++;
}
}
this.vocabSize = idx;
this.reverseVocab = Object.fromEntries(Object.entries(this.vocab).map(([k, v]) => [v, k]));
console.log(logSymbols.success, chalk.green(`Vocabulary built: ${this.vocabSize} tokens`));
console.log(chalk.dim(`📖 Sample vocab: ${Object.keys(this.vocab).slice(0, 15).join(", ")}`));
}
encode(text) {
return text.toLowerCase().split(/\s+/).filter(w => w).map(w => this.vocab[w] ?? this.vocab["<unk>"]);
}
decode(tokens) {
return tokens.map(t => this.reverseVocab[t] || "<unk>").join(" ");
}
}
class MathUtils {
static matmul(a, b) {
const m = a.length, n = a[0].length, p = b[0].length;
const result = Array(m).fill(0).map(() => Array(p).fill(0));
for (let i = 0; i < m; i++) {
for (let j = 0; j < p; j++) {
for (let k = 0; k < n; k++) {
result[i][j] += a[i][k] * b[k][j];
}
}
}
return result;
}
static matmul(a, b) {
const m = a.length, n = a[0].length, p = b[0].length;
const result = Array(m).fill(0).map(() => Array(p).fill(0));
for (let i = 0; i < m; i++) {
for (let j = 0; j < p; j++) {
for (let k = 0; k < n; k++) {
result[i][j] += a[i][k] * b[k][j];
}
}
}
return result;
}
static vecMatmul(vec, mat) {
const m = mat[0].length, n = vec.length;
const result = Array(m).fill(0);
for (let j = 0; j < m; j++) {
for (let i = 0; i < n; i++) {
result[j] += vec[i] * mat[i][j];
}
}
return result;
}
static outerProduct(vecA, vecB) {
const result = Array(vecA.length).fill(0).map(() => Array(vecB.length).fill(0));
for (let i = 0; i < vecA.length; i++) {
for (let j = 0; j < vecB.length; j++) {
result[i][j] = vecA[i] * vecB[j];
}
}
return result;
}
static transpose(matrix) {
if (!matrix || !matrix[0]) return [];
const rows = matrix.length;
const cols = matrix[0].length;
const result = Array(cols).fill(0).map(() => Array(rows).fill(0));
for (let i = 0; i < rows; i++) {
for (let j = 0; j < cols; j++) {
result[j][i] = matrix[i][j];
}
}
return result;
}
static add(a, b) { return a.map((val, i) => val + b[i]); }
static subtract(a, b) { return a.map((val, i) => val - b[i]); }
static scale(vec, s) { return vec.map(v => v * s); }
static relu(x) { return x.map(v => Math.max(0, v)); }
static reluDerivative(x) { return x.map(v => v > 0 ? 1 : 0); }
static softmax(logits) {
const maxLogit = Math.max(...logits);
const exp = logits.map(x => Math.exp(x - maxLogit));
const sum = exp.reduce((a, b) => a + b, 0);
return exp.map(x => x / sum);
}
static layerNorm(x) {
const mean = x.reduce((a, b) => a + b, 0) / x.length;
const variance = x.reduce((a, b) => a + (b - mean) ** 2, 0) / x.length;
const std = Math.sqrt(variance + 1e-5);
return x.map(val => (val - mean) / std);
}
static crossEntropy(probs, targetIdx) {
return -Math.log(Math.max(probs[targetIdx], 1e-10));
}
}
class MiniTransformer {
constructor(vocabSize, embedDim, hiddenDim, numLayers) {
this.vocabSize = vocabSize;
this.embedDim = embedDim;
this.hiddenDim = hiddenDim;
this.numLayers = numLayers;
this.embedding = this.randomMatrix(vocabSize, embedDim);
this.layers = Array(numLayers).fill(0).map(() => ({
attention: {
wq: this.randomMatrix(embedDim, embedDim),
wk: this.randomMatrix(embedDim, embedDim),
wv: this.randomMatrix(embedDim, embedDim),
wo: this.randomMatrix(embedDim, embedDim)
},
mlp: {
w1: this.randomMatrix(embedDim, hiddenDim),
b1: Array(hiddenDim).fill(0),
w2: this.randomMatrix(hiddenDim, embedDim),
b2: Array(embedDim).fill(0)
}
}));
this.outputWeights = this.randomMatrix(embedDim, vocabSize);
}
randomMatrix(rows, cols) {
const scale = Math.sqrt(2.0 / rows);
return Array(rows).fill(0).map(() =>
Array(cols).fill(0).map(() => (Math.random() - 0.5) * 2 * scale)
);
}
forward(tokens) {
this.cache = { tokens, layers: [] };
let x_sequence = tokens.map(t => this.embedding[t]);
for (const layer of this.layers) {
const layerCache = {};
const last_x = x_sequence[x_sequence.length - 1];
const q = MathUtils.vecMatmul(last_x, layer.attention.wq);
const k = MathUtils.vecMatmul(last_x, layer.attention.wk);
const v = MathUtils.vecMatmul(last_x, layer.attention.wv);
const attn_out = MathUtils.vecMatmul(v, layer.attention.wo);
let x = MathUtils.add(last_x, attn_out);
x = MathUtils.layerNorm(x);
layerCache.postAttn = [...x];
const mlp_hidden = MathUtils.add(MathUtils.vecMatmul(x, layer.mlp.w1), layer.mlp.b1);
layerCache.preRelu = [...mlp_hidden];
const mlp_activated = MathUtils.relu(mlp_hidden);
layerCache.postRelu = [...mlp_activated];
const mlp_out = MathUtils.add(MathUtils.vecMatmul(mlp_activated, layer.mlp.w2), layer.mlp.b2);
x = MathUtils.add(x, mlp_out);
x = MathUtils.layerNorm(x);
x_sequence[x_sequence.length - 1] = x;
this.cache.layers.push(layerCache);
}
const finalHidden = x_sequence[x_sequence.length - 1];
const logits = MathUtils.vecMatmul(finalHidden, this.outputWeights);
const probs = MathUtils.softmax(logits);
this.cache.finalHidden = finalHidden;
this.cache.probs = probs;
return probs;
}
backward(targetIdx, lr) {
let dLogits = [...this.cache.probs];
dLogits[targetIdx] -= 1;
const outputWeightsT = MathUtils.transpose(this.outputWeights);
const dFinalHidden = MathUtils.vecMatmul(dLogits, outputWeightsT);
const dOutputWeights = MathUtils.outerProduct(this.cache.finalHidden, dLogits);
for (let i = 0; i < this.embedDim; i++) {
for (let j = 0; j < this.vocabSize; j++) {
this.outputWeights[i][j] -= lr * dOutputWeights[i][j];
}
}
let dCurrent = dFinalHidden;
for (let l = this.numLayers - 1; l >= 0; l--) {
const layer = this.layers[l];
const cache = this.cache.layers[l];
const dMLP_out = dCurrent;
const w2_T = MathUtils.transpose(layer.mlp.w2);
const dHidden_activated = MathUtils.vecMatmul(dMLP_out, w2_T);
const dW2 = MathUtils.outerProduct(cache.postRelu, dMLP_out);
const dB2 = dMLP_out;
const dHidden_preRelu = dHidden_activated.map((g, i) => g * (cache.preRelu[i] > 0 ? 1 : 0));
const w1_T = MathUtils.transpose(layer.mlp.w1);
dCurrent = MathUtils.vecMatmul(dHidden_preRelu, w1_T);
const dW1 = MathUtils.outerProduct(cache.postAttn, dHidden_preRelu);
const dB1 = dHidden_preRelu;
for(let i=0; i<this.hiddenDim; i++) for(let j=0; j<this.embedDim; j++) layer.mlp.w2[i][j] -= lr * dW2[i][j];
for(let j=0; j<this.embedDim; j++) layer.mlp.b2[j] -= lr * dB2[j];
for(let i=0; i<this.embedDim; i++) for(let j=0; j<this.hiddenDim; j++) layer.mlp.w1[i][j] -= lr * dW1[i][j];
for(let j=0; j<this.hiddenDim; j++) layer.mlp.b1[j] -= lr * dB1[j];
}
const dEmbedding = dCurrent;
const lastTokenId = this.cache.tokens[this.cache.tokens.length - 1];
for (let i = 0; i < this.embedDim; i++) {
this.embedding[lastTokenId][i] -= lr * dEmbedding[i];
}
}
saveWeights(filename) {
const weights = {
vocabSize: this.vocabSize, embedDim: this.embedDim, hiddenDim: this.hiddenDim,
numLayers: this.numLayers, embedding: this.embedding, layers: this.layers,
outputWeights: this.outputWeights
};
fs.writeFileSync(filename, JSON.stringify(weights));
console.log(logSymbols.success, chalk.green(`Weights saved to ${chalk.bold(filename)}`));
}
}
function train(config) {
console.log("\n" + chalk.bold.yellow("🚀 Starting Transformer Training...") + "\n" + "=".repeat(60));
const texts = loadTinyChatDataset(config.datasetPath);
const tokenizer = new Tokenizer();
tokenizer.buildVocab(texts, 2);
console.log(logSymbols.info, chalk.blue(`\nCreating training data (up to ${config.maxSamples} samples)...`));
const trainingPairs = [];
for (const text of texts) {
const tokens = tokenizer.encode(text);
for (let i = 0; i < tokens.length - 1; i++) {
trainingPairs.push({
input: tokens.slice(Math.max(0, i - config.contextWindow + 1), i + 1),
target: tokens[i + 1]
});
if (trainingPairs.length >= config.maxSamples) break;
}
if (trainingPairs.length >= config.maxSamples) break;
}
console.log(logSymbols.success, chalk.green(`Created ${trainingPairs.length} training examples.`));
const model = new MiniTransformer(tokenizer.vocabSize, config.embedDim, config.hiddenDim, config.numLayers);
console.log(chalk.cyan.bold("\n🧠 Model Architecture:"));
console.log(chalk.cyan(` - Vocab Size: ${tokenizer.vocabSize}`));
console.log(chalk.cyan(` - Embedding Dim: ${model.embedDim}`));
console.log(chalk.cyan(` - Hidden Dim: ${model.hiddenDim}`));
console.log(chalk.cyan(` - Layers: ${model.numLayers}\n`));
console.log("=".repeat(60) + chalk.bold.yellow("\n📈 Training Progress:\n"));
for (let epoch = 0; epoch < config.epochs; epoch++) {
let totalLoss = 0;
const shuffled = trainingPairs.sort(() => Math.random() - 0.5);
for (let i = 0; i < shuffled.length; i++) {
const { input, target } = shuffled[i];
const probs = model.forward(input);
totalLoss += MathUtils.crossEntropy(probs, target);
model.backward(target, config.learningRate);
if ((i + 1) % 1000 === 0) {
process.stdout.write(chalk.dim(`\rEpoch ${epoch + 1}/${config.epochs} | Batch ${i+1}/${shuffled.length}`));
}
}
const avgLoss = totalLoss / trainingPairs.length;
process.stdout.clearLine(0);
process.stdout.cursorTo(0);
console.log(
chalk.magenta(`Epoch ${(epoch + 1).toString().padStart(2)}/${config.epochs}`) +
chalk.gray(` | `) +
chalk.green(`Loss: ${avgLoss.toFixed(4)}`)
);
}
console.log("=".repeat(60));
model.saveWeights(config.weightsFile);
fs.writeFileSync(config.tokenizerFile, JSON.stringify({ vocab: tokenizer.vocab, vocabSize: tokenizer.vocabSize }, null, 2));
console.log(logSymbols.success, chalk.green(`Tokenizer saved to ${chalk.bold(config.tokenizerFile)}`));
console.log("\n" + "=".repeat(60) + chalk.bold.yellow("\n🎯 TESTING GENERATION\n"));
const testPrompts = ["What do you", "I like to", "Do you have"];
testPrompts.forEach(prompt => {
let tokens = tokenizer.encode(prompt);
let generated = [];
for (let i = 0; i < 10; i++) {
const probs = model.forward(tokens.slice(-config.contextWindow));
const nextToken = probs.indexOf(Math.max(...probs));
if (nextToken === tokenizer.vocab["<eos>"]) break;
generated.push(nextToken);
tokens.push(nextToken);
}
console.log(chalk.blue(`Prompt: "${prompt}"`));
console.log(chalk.green(`Output: "${tokenizer.decode([...tokenizer.encode(prompt), ...generated])}"\n`));
});
console.log("=".repeat(60) + "\n" + logSymbols.success, chalk.bold.bgGreen.black(" Training complete! "));
}
if (!fs.existsSync(TRAINING_CONFIG.datasetPath)) {
console.error(logSymbols.error, chalk.red(`Error: Dataset file not found at ${chalk.bold(TRAINING_CONFIG.datasetPath)}!`));
process.exit(1);
}
train(TRAINING_CONFIG); |