335 NLP Notebooks That Refuse to Upgrade to TensorFlow 2

What it does

This repository collects 335 Jupyter notebooks implementing machine learning and deep learning models across roughly two dozen NLP tasks, including chatbots, neural machine translation, dependency parsing, and OCR. Each notebook targets the narrow TensorFlow window between 1.13 and 2.0, and the author’s stated goal is to simplify research-grade code that is “not really beginner friendly.” Where the author did not build a model from scratch, the README admits to copying, pasting, and fixing third-party code for deprecated API issues.

The interesting bit

The value is in the side-by-side volume: you can browse 54 chatbot variants or 53 neural machine translation architectures, ranging from basic LSTM seq2seq to stacked Bahdanau-Luong attention with beam decoders. It also acts as a time capsule of pre- and early-Transformer techniques, with some notebooks implementing papers that were not yet released at the time of writing.

Key highlights

• 335 notebooks spanning about 25 categories, from abstractive summarization to old-to-young vocoding.
• Explicitly simplified implementations aimed at learners, with test accuracy numbers provided for quick comparison (for example, a pointer-generator summarizer hitting 16.51%).
• Adapted third-party code is flagged and linked, making the repo a curated index rather than a from-scratch rewrite.
• Strictly bound to TensorFlow 1.x graph-mode execution; no TensorFlow 2 support is included.

Caveats

• Locked to a legacy TensorFlow version, so modern eager execution, Keras layers, and current API patterns are absent.
• Some reported accuracies are based on limited epochs or narrow datasets, and the README does not distinguish between proof-of-concept baselines and competitive results.
• The author openly notes that some entries are copy-paste-fix jobs, so production readiness is not the focus.

Verdict

A useful reference if you are learning how classic NLP architectures were wired in the TF 1.x era or need a quick prototype of an attention variant. Avoid if you require modern TensorFlow 2, PyTorch, or battle-hardened production code.