google-research/electra

Health v2 · maintenance only

Flags: no_releases archived

Adoption not part of the score

2367 stars · 347 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

ELECTRA is a research library from Google for self-supervised pre-training of transformer text encoders using a discriminator-based objective that replaces masked tokens with generated fakes. It includes pre-training and fine-tuning code plus released small/base/large English models, and the Electric energy-based cloze variant.

Use cases

• pre-train a small transformer language model on a single GPU
• fine-tune a pretrained encoder for text classification on GLUE
• build a question answering model on SQuAD
• train a sequence tagging model for text chunking
• score text with pseudo-likelihoods to re-rank speech recognition or machine translation outputs

When to choose

• you want strong encoder pre-training with limited compute
• you need a small efficient BERT alternative for classification, QA, or tagging
• you are reproducing or extending ELECTRA/Electric research

When to avoid

• you need a maintained production NLP framework or recent model releases
• you want multilingual or generative (decoder) models
• you prefer PyTorch over TensorFlow

Facets

library · maturity maintenance

machine-learning deep-learning nlp llm-training deep-learning machine-learning python transformers pre-training language-models tensorflow research-code discriminative-pretraining natural-language-processing gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="google-research/electra")