What it does This repo implements a vanilla RNN language model using only NumPy. It walks through forward propagation, cross-entropy loss, and SGD training, with all the gate-level operations—multiply, add, tanh, softmax—written out explicitly as classes with forward and backward methods.

The interesting bit The author frames Backpropagation Through Time as ordinary backprop on an unrolled computation graph, then simplifies it by grouping operations into composite units. It is a pedagogical trick: instead of hand-deriving scary partial derivatives, you let the graph structure do the bookkeeping.

Key highlights

• Pure NumPy implementation with explicit MultiplyGate, AddGate, Tanh, and Softmax primitives
• BPTT truncated to a configurable window (bptt_truncate=4) to keep the unrolled graph manageable
• Xavier-style initialization scaled by incoming connection count
• Code mirrors a companion “neural network from scratch” repo, so the progression from feedforward to recurrent is intentional

Caveats

• The README trails off mid-sentence during the SGD section (evaluate_loss_ is cut off), so the full training loop is incomplete in the source
• No performance numbers, dataset details, or convergence plots are shown; it is strictly a tutorial skeleton

Verdict Grab this if you are teaching RNNs or finally want to see BPTT without PyTorch’s autograd hiding the wires. Skip it if you need a production language model or GPU acceleration.