Abhisheksvnit/ElectraAI-Dataset-v4

ElectraAI Dataset v4

1M+ record ECE/VLSI/Analog/Embedded dataset for LLM and vision model fine-tuning.

Statistics

Stream	Records	Description
A: NGSpice Netlists	100,000	SPICE netlists + analyses
B: Labeled Circuit Images	100,000	PNG + JSON (schemdraw)
C: QA Pairs (ShareGPT)	700,000	Multi-turn ECE instruction
D: YOLO Annotations	100,000	PNG + YOLO TXT (20 classes)
Total	1,000,000

Repo layout (sharded directories)

Images and labels are sharded into subfolders of ≤5,000 files each (e.g. stream_B_circuit_images/images/shard_0000/, shard_0001/, ...) to stay under HuggingFace's 10,000-files-per-directory repo limit.

Topology Coverage (500+ topologies across 10 domains)

Domain	Topologies	QA Records
Analog Electronics	100	95,000
Analog VLSI	100	85,000
Network Theory	100	95,000
Digital VLSI + STA	60	80,000
Embedded Systems	50	70,000
Signal Processing	50	60,000
Communication	50	60,000
Microprocessors	50	50,000
Power Electronics	50	65,000
AI + VLSI	40	40,000

YOLO Class Definitions (Stream D)

20 component classes: resistor, capacitor, inductor, diode, zener_diode, bjt_npn, bjt_pnp, nmos, pmos, opamp, voltage_source, current_source, ground, wire_junction, transformer, switch, voltage_probe, current_probe, ic_block, logic_gate.

Usage

from datasets import load_dataset
# QA pairs for LLM fine-tuning
ds = load_dataset("Abhisheksvnit/ElectraAI-Dataset-v4", data_files="stream_C_qa_pairs/*.jsonl")

Generated by ElectraAI v4 pipeline (Kaggle T4×2/P100, zero cost).