0xShashi/tech-talks-segments

Awesome Tech Talks Dataset

A curated dataset of 2,200+ technical sessions, workshops, and keynotes from official engineering organizations including Google, Microsoft, OpenAI, Anthropic, and Cursor. The dataset includes video metadata, structured topic classifications, cleaned transcripts, and 42,000+ segmented text chunks designed for Retrieval-Augmented Generation (RAG), vector search, and language model evaluation.

Dataset Summary

Attribute	Value
Total Unique Videos	2,271
Total Transcript Segments	42,583
Primary Language	English (en)
Source Platforms	YouTube official channels, unlisted event replays
Supported Formats	Parquet, JSONL
License	MIT (Dataset schema, curation scripts, and annotations)

Dataset Files

The repository provides two levels of granularity:

File	Records	Description
tech_talks_segments.parquet	42,583	Chunked and segmented transcripts with timestamps and section headings for RAG pipelines.
tech_talks_unified.parquet	2,271	Video-level records with metadata, topic lists, and complete full-text transcripts.

Dataset Structure

Segment Level Schema (tech_talks_segments.parquet)

Field	Type	Description
segment_id	string	Unique identifier for the segment (video_id + segment index)
video_id	string	Canonical YouTube video identifier
title	string	Title of the talk or session
company	string	Organization that hosted or published the session
channel	string	Channel name where the video was originally uploaded
video_type	string	Visibility status (public or unlisted)
upload_date	string	Original publication date (YYYY-MM-DD)
duration_seconds	float	Total duration of the full video in seconds
url	string	Source URL to the original session
description	string	Original video description text
topics	list[string]	List of standardized topic tags assigned to the session
qc_flags	string	Quality control or validation notices, if any
segment_index	int64	Positional index of the segment within the video
total_segments	int64	Total count of segments for this video
heading	string	Extracted section heading or topic transition title
heading_confidence	string	Confidence level of the heading extraction (high or low)
start_time	string	Timestamp offset where the segment begins (HH:MM:SS)
text	string	Cleaned, normalized transcript text for the segment
word_count	int64	Total word count of the segment

Video Level Schema (tech_talks_unified.parquet)

Field	Type	Description
video_id	string	Canonical YouTube video identifier
company	string	Host organization
channel	string	Source channel name
video_type	string	Visibility status (public or unlisted)
title	string	Full video title
upload_date	string	Date published
shareable_url	string	Direct link to the video
duration_seconds	float	Total video length in seconds
description	string	Video description
topics	list[string]	Taxonomy topic tags
segments	list[dict]	Array of segment dictionaries
full_transcript	string	Complete concatenated transcript text

Organization Breakdown

Organization	Channels Covered	Public Sessions	Unlisted Sessions	Total Sessions
Microsoft	1	1,694	0	1,694
Google	5	184	71	255
Anthropic	3	68	72	140
OpenAI	2	74	48	122
Cursor	1	64	1	65
Total	12	2,082	189	2,271

Topic Taxonomy

Sessions are tagged with one to three standard categories from the following taxonomy:

Category Group	Example Topics
Artificial Intelligence	AI Agents, LLM Fundamentals, Prompt Engineering, Evaluation, Fine-Tuning
Cloud and Infrastructure	Kubernetes, Distributed Systems, Cloud Architecture, Serverless, DevOps
Software Engineering	API Design, System Architecture, Performance Optimization, Testing
Developer Tooling	AI Coding Assistants, IDEs, CI/CD, Compilers, Developer Productivity
Data and Analytics	Data Engineering, Vector Databases, Retrieval Systems, Analytics Pipelines

How to Use

Loading via Datasets Library

from datasets import load_dataset

# Load segmented records (recommended for RAG and chunk-level search)
dataset = load_dataset("your-username/tech-talks-segments", split="train")

# Preview first record
print(dataset[0])

Loading via Pandas directly

import pandas as pd

df = pd.read_parquet("hf://datasets/your-username/tech-talks-segments/data/tech_talks_segments.parquet")
print(df[["title", "company", "heading", "word_count"]].head())

Intended Use Cases

Use Case	Description
Retrieval-Augmented Generation (RAG)	Pre-chunked segment text and timestamps allow building searchable knowledge bases for engineering queries.
Technical QA Evaluation	Benchmark LLM question-answering accuracy against authoritative technical presentations.
Trend Analysis	Analyze shifts in engineering focus across major tech organizations over time.
Domain Adaptation	Fine-tune language models or embeddings on clean technical speech and developer vocabulary.

Data Collection and Quality Control

1. Source Selection: Video lists were gathered strictly from verified organization channels, engineering conference playlists, and unlisted developer workshop archives.
2. Transcript Processing: Auto-generated and manual transcripts were extracted, cleaned of repetitive filler phrases, normalized, and aligned with timestamps.
3. Segmentation: Transcripts were split into logical segments based on speaker shifts, visual slide boundaries, and topical headings.
4. Quality Filtering: Entries missing valid transcripts or containing corrupted metadata were cataloged with quality flags in the QC audit log.

Licensing and Attribution

Component	License	Notes
Dataset Structure and Metadata	MIT License	Covers schemas, segmentation logic, tags, and curation pipeline.
Original Video and Audio Content	Third-Party Copyright	All video content, titles, and trademarks remain the intellectual property of their respective owners (Google, Microsoft, Anthropic, OpenAI, Cursor).

This dataset is distributed for educational, research, and non-commercial technical evaluation purposes. Inclusion of sessions does not imply endorsement, sponsorship, or affiliation with the featured companies.

Project Links

Resource	Link
GitHub Repository	https://github.com/0x-Shashi/awesome-tech-talks
Catalog by Company	Available in GitHub repository under catalog/by-company.md
Catalog by Topic	Available in GitHub repository under catalog/by-topic/