YouTube Synthetic Dataset Workflow

Generate synthetic Q&A pairs from YouTube videos, create snapshots, and launch fine-tuning.

Overview

This walkthrough demonstrates how to build a stock analysis and investment tips extraction model from YouTube financial content. You’ll learn how to:

• Process 5 YouTube videos about stocks and investing
• Use question/answer templates to extract simple, structured JSON output
• Generate training data that captures stock tickers, actionable tips, and risks mentioned
• Create train/validation splits and get model recommendations
• Launch fine-tuning to create a specialized financial intelligence model

Why This Matters: Financial YouTube channels contain valuable insights, but this information is locked in video format. This workflow extracts structured intelligence from videos and transforms it into a model that can analyze financial content and output organized data.

What You’ll Build: A model that analyzes financial video transcripts and outputs clean JSON with:

• stocks_mentioned: Array of ticker symbols (e.g., [“AAPL”, “TSLA”])
• investment_tips: List of actionable advice from the video
• risks_mentioned: Warnings or cautionary statements discussed
• video_topic: Brief description of the main topic

Export your Prem API key as API_KEY before running any script. The example uses 5 curated financial YouTube videos. You can modify the YOUTUBE_URLS array to use your own videos.

1. Setup: Define your YouTube URLs

   TypeScript

   const API_KEY = process.env.API_KEY;
   
   // Define the YouTube videos you want to analyze
   const YOUTUBE_URLS = [
     'https://www.youtube.com/watch?v=JH-k5f4Yclc',
     'https://www.youtube.com/watch?v=YEWhxcpMS1c',
     'https://www.youtube.com/watch?v=cb8up3HVXis',
     'https://www.youtube.com/watch?v=26xatIiMv88',
     'https://www.youtube.com/watch?v=-Da3gUdzCvs'
   ];
   

   Define the YouTube video URLs you want to process. You can replace these with any financial or educational videos.

2. Create project and generate synthetic dataset

   TypeScript

   const res = await fetch('https://studio.premai.io/api/v1/public/projects/create', {
     method: 'POST',
     headers: {
       'Authorization': `Bearer ${API_KEY}`,
       'Content-Type': 'application/json'
     },
     body: JSON.stringify({ name: 'Stock Analysis Project', goal: 'Extract investment insights from financial videos' })
   });
   if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
   const { project_id } = await res.json();
   
   const formData = new FormData();
   formData.append('project_id', project_id);
   formData.append('name', 'Financial YouTube Dataset');
   
   // Add YouTube URLs
   YOUTUBE_URLS.forEach((url: string, index: number) => {
     formData.append(`youtube_urls[${index}]`, url);
   });
   
   formData.append('pairs_to_generate', '50');
   formData.append('pair_type', 'qa');
   formData.append('temperature', '0.3');
   
   // Add rules and constraints
   formData.append('rules[]', 'stocks_mentioned: List all stock ticker symbols mentioned (e.g., AAPL, TSLA, NVDA)');
   formData.append('rules[]', 'investment_tips: Extract 3-5 specific, actionable pieces of advice from the video');
   formData.append('rules[]', 'risks_mentioned: List any warnings, risks, or cautionary statements discussed');
   formData.append('rules[]', 'video_topic: Write a short phrase describing the main topic of the video');
   formData.append('rules[]', 'Only output valid JSON with no additional text before or after');
   formData.append('rules[]', 'If a field has no relevant information, use an empty array [] or empty string ""');
   formData.append('rules[]', 'Use exact quotes or close paraphrases from the video content');
   formData.append('rules[]', 'Do not invent or infer information not explicitly stated');
   
   // Define question format
   const questionFormat = `Extract investment information from the following video transcript:
   
   {VIDEO_TRANSCRIPT}
   
   Provide the output in this JSON format:
   {
   "stocks_mentioned": ["TICKER1", "TICKER2"],
   "investment_tips": ["tip 1", "tip 2", "tip 3"],
   "risks_mentioned": ["risk 1", "risk 2"],
   "video_topic": "brief description of main topic"
   }`;
   formData.append('question_format', questionFormat);
   
   // Define answer format
   const answerFormat = `{
   "stocks_mentioned": ["<TICKER_SYMBOL_1>", "<TICKER_SYMBOL_2>"],
   "investment_tips": ["<specific_tip_1>", "<specific_tip_2>", "<specific_tip_3>"],
   "risks_mentioned": ["<risk_or_warning_1>", "<risk_or_warning_2>"],
   "video_topic": "<main_topic_of_video>"
   }`;
   formData.append('answer_format', answerFormat);
   
   const res2 = await fetch('https://studio.premai.io/api/v1/public/datasets/create-synthetic', {
     method: 'POST',
     headers: { 'Authorization': `Bearer ${API_KEY}` },
     body: formData
   });
   if (!res2.ok) throw new Error(`${res2.status}: ${await res2.text()}`);
   const { dataset_id } = await res2.json();
   

   Create a project, then submit 5 YouTube URLs about stocks and investing to generate synthetic Q&A pairs. The rules clearly define how each JSON field should be populated (stocks as tickers, tips as actionable advice, risks as warnings, topic as brief description). The generation can take 5-10 minutes.

3. Wait for dataset generation

   TypeScript

   let dataset;
   let checks = 0;
   do {
     await sleep(5000);
     const res = await fetch(`https://studio.premai.io/api/v1/public/datasets/${dataset_id}`, {
       headers: { 'Authorization': `Bearer ${API_KEY}` }
     });
     if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
     dataset = await res.json();
     if (checks++ % 6 === 0) {
       console.log(`Status: ${dataset.status}, ${dataset.datapoints_count} datapoints`);
     }
   } while (dataset.status === 'processing');
   

   Poll the dataset status every 5 seconds. Log progress every 30 seconds to avoid spam.

4. Create snapshot and get recommendations

   TypeScript

   const res = await fetch('https://studio.premai.io/api/v1/public/snapshots/create', {
     method: 'POST',
     headers: {
       'Authorization': `Bearer ${API_KEY}`,
       'Content-Type': 'application/json'
     },
     body: JSON.stringify({ dataset_id, split_percentage: 80 })
   });
   if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
   const { snapshot_id } = await res.json();
   
   const res2 = await fetch('https://studio.premai.io/api/v1/public/recommendations/generate', {
     method: 'POST',
     headers: {
       'Authorization': `Bearer ${API_KEY}`,
       'Content-Type': 'application/json'
     },
     body: JSON.stringify({ snapshot_id })
   });
   if (!res2.ok) throw new Error(`${res2.status}: ${await res2.text()}`);
   
   let recs;
   do {
     await sleep(5000);
     const res3 = await fetch(`https://studio.premai.io/api/v1/public/recommendations/${snapshot_id}`, {
       headers: { 'Authorization': `Bearer ${API_KEY}` }
     });
     if (!res3.ok) throw new Error(`${res3.status}: ${await res3.text()}`);
     recs = await res3.json();
   } while (recs.status === 'processing');
   

   Create a snapshot with 80/20 train/validation split, then request model recommendations.

5. Launch fine-tuning job

   TypeScript

   const experiments = recs.recommended_experiments
     .filter((e: any) => e.recommended)
     .map(({ recommended, reason_for_recommendation, ...experiment }: any) => experiment);
   
   const res = await fetch('https://studio.premai.io/api/v1/public/finetuning/create', {
     method: 'POST',
     headers: {
       'Authorization': `Bearer ${API_KEY}`,
       'Content-Type': 'application/json'
     },
     body: JSON.stringify({ snapshot_id, name: 'YouTube Model', experiments })
   });
   if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
   const { job_id } = await res.json();
   

   Filter recommended experiments and use them directly to launch the fine-tuning job.

6. Monitor job progress

   TypeScript

   for (let i = 0; i < 30; i++) {
     await sleep(10000);
     const res = await fetch(`https://studio.premai.io/api/v1/public/finetuning/${job_id}`, {
       headers: { 'Authorization': `Bearer ${API_KEY}` }
     });
     if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
     const job = await res.json();
     console.log(`Status: ${job.status}`);
     job.experiments.forEach((e: any) => {
       console.log(`  - Exp #${e.experiment_number}: ${e.status} ${e.model_id || ''}`);
     });
     if (job.status !== 'processing') break;
   }
   

   Poll job status every 10 seconds. Each experiment shows its progress and final model ID.

Full Example

TypeScript

#!/usr/bin/env bun

/**
 * Example 2: YouTube synthetic dataset workflow
 * 1. Create project → 2. Generate synthetic data from YouTube → 3. Create snapshot → 4. Get recommendations → 5. Run finetuning
 */

const API_KEY = process.env.API_KEY;
const YOUTUBE_URLS = [
	'https://www.youtube.com/watch?v=JH-k5f4Yclc',
	'https://www.youtube.com/watch?v=YEWhxcpMS1c',
	'https://www.youtube.com/watch?v=cb8up3HVXis',
	'https://www.youtube.com/watch?v=26xatIiMv88',
	'https://www.youtube.com/watch?v=-Da3gUdzCvs'
];

if (!API_KEY) {
	console.error('Error: API_KEY environment variable is required');
	console.error('Please create a .env file based on .env.example');
	process.exit(1);
}

function sleep(ms: number) {
	return new Promise((r) => setTimeout(r, ms));
}

async function main() {
	console.log('\n=== YouTube Synthetic Workflow ===\n');

	// 1. Create project
	console.log('1. Creating project...');
	const res1 = await fetch('https://studio.premai.io/api/v1/public/projects/create', {
		method: 'POST',
		headers: {
			'Authorization': `Bearer ${API_KEY}`,
			'Content-Type': 'application/json'
		},
		body: JSON.stringify({ name: 'Stock Analysis Project', goal: 'Extract investment insights from financial videos' }),
	});
	if (!res1.ok) throw new Error(`${res1.status}: ${await res1.text()}`);
	const { project_id } = await res1.json();
	console.log(`   ✓ Project: ${project_id}\n`);

	// 2. Generate synthetic dataset
	console.log('2. Generating synthetic dataset from YouTube...');
	console.log(`   URLs: ${YOUTUBE_URLS.length} financial videos`);
	const formData = new FormData();
	formData.append('project_id', project_id);
	formData.append('name', 'Financial YouTube Dataset');

	// Add multiple YouTube URLs
	YOUTUBE_URLS.forEach((url, index) => {
		formData.append(`youtube_urls[${index}]`, url);
	});

	formData.append('pairs_to_generate', '50');
	formData.append('pair_type', 'qa');
	formData.append('temperature', '0.3');

	// Add rules and constraints
	formData.append('rules[]', 'stocks_mentioned: List all stock ticker symbols mentioned (e.g., AAPL, TSLA, NVDA)');
	formData.append('rules[]', 'investment_tips: Extract 3-5 specific, actionable pieces of advice from the video');
	formData.append('rules[]', 'risks_mentioned: List any warnings, risks, or cautionary statements discussed');
	formData.append('rules[]', 'video_topic: Write a short phrase describing the main topic of the video');
	formData.append('rules[]', 'Only output valid JSON with no additional text before or after');
	formData.append('rules[]', 'If a field has no relevant information, use an empty array [] or empty string ""');
	formData.append('rules[]', 'Use exact quotes or close paraphrases from the video content');
	formData.append('rules[]', 'Do not invent or infer information not explicitly stated');

	// Define question format
	const questionFormat = `Extract investment information from the following video transcript:

{VIDEO_TRANSCRIPT}

Provide the output in this JSON format:
{
  "stocks_mentioned": ["TICKER1", "TICKER2"],
  "investment_tips": ["tip 1", "tip 2", "tip 3"],
  "risks_mentioned": ["risk 1", "risk 2"],
  "video_topic": "brief description of main topic"
}`;
	formData.append('question_format', questionFormat);

	// Define answer format
	const answerFormat = `{
  "stocks_mentioned": ["<TICKER_SYMBOL_1>", "<TICKER_SYMBOL_2>"],
  "investment_tips": ["<specific_tip_1>", "<specific_tip_2>", "<specific_tip_3>"],
  "risks_mentioned": ["<risk_or_warning_1>", "<risk_or_warning_2>"],
  "video_topic": "<main_topic_of_video>"
}`;
	formData.append('answer_format', answerFormat);

	const res2 = await fetch('https://studio.premai.io/api/v1/public/datasets/create-synthetic', {
		method: 'POST',
		headers: { 'Authorization': `Bearer ${API_KEY}` },
		body: formData,
	});
	if (!res2.ok) throw new Error(`${res2.status}: ${await res2.text()}`);
	const { dataset_id } = await res2.json();
	console.log(`   ✓ Dataset: ${dataset_id}`);

	// Wait for dataset (can take several minutes)
	console.log('   Waiting for generation (may take 5-10 minutes)...');
	let dataset;
	let checks = 0;
	do {
		await sleep(5000);
		const res = await fetch(`https://studio.premai.io/api/v1/public/datasets/${dataset_id}`, {
			headers: { 'Authorization': `Bearer ${API_KEY}` }
		});
		if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
		dataset = await res.json();
		if (checks++ % 6 === 0) {
			console.log(`   Status: ${dataset.status}, ${dataset.datapoints_count} datapoints`);
		}
	} while (dataset.status === 'processing');
	console.log(`   ✓ Ready: ${dataset.datapoints_count} datapoints\n`);

	// 3. Create snapshot
	console.log('3. Creating snapshot...');
	const res3 = await fetch('https://studio.premai.io/api/v1/public/snapshots/create', {
		method: 'POST',
		headers: {
			'Authorization': `Bearer ${API_KEY}`,
			'Content-Type': 'application/json'
		},
		body: JSON.stringify({ dataset_id, split_percentage: 80 }),
	});
	if (!res3.ok) throw new Error(`${res3.status}: ${await res3.text()}`);
	const { snapshot_id } = await res3.json();
	console.log(`   ✓ Snapshot: ${snapshot_id}\n`);

	// 4. Generate recommendations
	console.log('4. Generating recommendations...');
	const res4 = await fetch('https://studio.premai.io/api/v1/public/recommendations/generate', {
		method: 'POST',
		headers: {
			'Authorization': `Bearer ${API_KEY}`,
			'Content-Type': 'application/json'
		},
		body: JSON.stringify({ snapshot_id }),
	});
	if (!res4.ok) throw new Error(`${res4.status}: ${await res4.text()}`);

	let recs;
	do {
		await sleep(5000);
		const res = await fetch(`https://studio.premai.io/api/v1/public/recommendations/${snapshot_id}`, {
			headers: { 'Authorization': `Bearer ${API_KEY}` }
		});
		if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
		recs = await res.json();
	} while (recs.status === 'processing');

	console.log(`   ✓ Recommended experiments:`);
	const recommendedCount = recs.recommended_experiments.filter((e: any) => e.recommended).length;
	console.log(`   Total experiments: ${recs.recommended_experiments.length}, Recommended: ${recommendedCount}`);
	recs.recommended_experiments.forEach((e: any) => {
		if (e.recommended) console.log(`     - ${e.base_model_id} (LoRA: ${e.lora})`);
	});
	console.log();

	// 5. Create finetuning job
	console.log('5. Creating finetuning job...');
	const experiments = recs.recommended_experiments
		.filter((e: any) => e.recommended)
		.map(({ recommended, reason_for_recommendation, ...experiment }: any) => experiment);

	if (experiments.length === 0) {
		console.error('\n✗ Error: No recommended experiments found. Cannot create finetuning job.');
		process.exit(1);
	}

	const res5 = await fetch('https://studio.premai.io/api/v1/public/finetuning/create', {
		method: 'POST',
		headers: {
			'Authorization': `Bearer ${API_KEY}`,
			'Content-Type': 'application/json'
		},
		body: JSON.stringify({ snapshot_id, name: 'YouTube Model', experiments }),
	});
	if (!res5.ok) throw new Error(`${res5.status}: ${await res5.text()}`);
	const { job_id } = await res5.json();
	console.log(`   ✓ Job: ${job_id}\n`);

	// 6. Monitor (5 minutes max)
	console.log('6. Monitoring job...');
	for (let i = 0; i < 30; i++) {
		await sleep(10000);
		const res = await fetch(`https://studio.premai.io/api/v1/public/finetuning/${job_id}`, {
			headers: { 'Authorization': `Bearer ${API_KEY}` }
		});
		if (!res.ok) throw new Error(`${res.status}: ${await res.text()}`);
		const job = await res.json();
		console.log(`   Status: ${job.status}`);
		job.experiments.forEach((e: any) => {
			console.log(`     - Exp #${e.experiment_number}: ${e.status} ${e.model_id || ''}`);
		});
		if (job.status !== 'processing') break;
	}

	console.log('\n✓ Done!\n');
}

main().catch((err) => {
	console.error('\n✗ Error:', err.message);
	process.exit(1);
});