What it does This is a full-semester crash course in deep learning for NLP, originally taught at ABBYY. It covers the standard arc from word embeddings through RNNs, CNNs, seq2seq, attention, and transformers, all the way to dialogue systems and pretrained models like BERT and ELMo. Every week is a Jupyter notebook you can open directly in Colab and run.

The interesting bit The course doesn’t just explain architectures—it makes you implement them. Week 3 has you build CBoW, skip-gram, and negative sampling from scratch in PyTorch. Week 7 asks you to generate Russian troll tweets with a character-level language model. The assignments are deliberately playful, which keeps the slog of implementing yet another RNN tolerable.

Key highlights

• 14 weekly notebooks, each with Russian originals and gradually-updated English versions
• Covers both classics (Goldberg’s textbook is the suggested reading) and then-recent advances like Pointer-Generator Networks and DSSMs
• Mixes toy problems (surname detection, poetry generation) with real competition data (Toxic Comment Classification, SQuAD)
• All notebooks are Colab-ready with one-click badges
• Final week explicitly covers pretrained models—useful if you’re trying to bridge from “I built an LSTM once” to “I fine-tuned BERT”

Caveats

• The author notes English translations are still in progress; some weeks only have Russian versions
• Materials are from an earlier era of NLP (pre-LLM frenzy), so don’t expect instruction tuning or RLHF
• No solutions or autograders visible—this is raw course material, not a polished MOOC

Verdict Good fit if you learn by breaking things in notebooks and want a structured path through pre-transformer NLP fundamentals. Skip it if you need cutting-edge LLM training or a fully-maintained, error-free experience.