character-level

Multi-layer Recurrent Neural Networks (LSTM, GRU, RNN) for character-level language models in Torch. Based on code of https://github.com/karpathy/char-rnn. Support Chinese and other things.

Code accompanying Incorporating Chinese Characters of Words for Lexical Sememe Prediction (ACL2018) https://arxiv.org/abs/1806.06349

Multi lingual character based named entity recognizer

Implementation of the Character-level Intra Attention Network (CIAN) for Natural Language Inference (NLI) upon SNLI and MultiNLI corpus

AI name parsing. Predict first or last name using a DL model.

LittleLM: A tiny character-level n-gram language model for local corpus building and collaborative experimentation.

a 5M parameter solution to a problem you could solve by counting on your fingers

🤖 From-scratch GPT in PyTorch — bigram, single-head & multi-head attention • FastAPI /generate endpoint • React + Vite frontend • Character-level tokenizer • CUDA support

Character-level vanilla RNN in PyTorch that predicts a name's language of origin, letter by letter. Includes training, confusion matrix evaluation, and inference — plus a real example of dataset distribution shift when tested on modern names.

In this project, I worked with a small corpus consisting of simple sentences. I tokenized the words using n-grams from the NLTK library and performed word-level and character-level one-hot encoding. Additionally, I utilized the Keras Tokenizer to tokenize the sentences and implemented word embedding using the Embedding layer. For sentiment analysis

generative model for names

Um GPT em nível de caractere escrito do zero em Java puro, sem bibliotecas de machine learning. Três níveis: bigrama por contagem, MLP com backpropagation manual e um Transformer com atenção causal.

A small experimental model based on SP-JEPA. Character-level text prediction in vector space.

build your own GPT, one letter at a time - a character-level GPT trained from scratch on public-domain books: trainer, corpus pipeline, benchmark harness, and 5 trained models

Learning experiment: training a tiny character-level GPT on Shakespeare, following Karpathy’s nanoGPT. My notes and runs.

TapeLM: facts as fingerprints on character ink — not token-id memory, not chunk RAG. One frozen curve encoder for generation and structured slot memory (write, bind, hop, resolve). Noisy recall, lexicon calibration, one-shot edits, clean unlearning — vs fair GPT/RAG; reproducible JSON benchmarks.

TinyTalker is a small GPT model based on a character level tokenization system built from scratch. You can prompt it or train it on your own data super easily. It is also super small and can be ran on laptops.

A tiny character-level GPT written from scratch in PyTorch (nanoGPT-style), trained on human proteins, E. coli DNA or Shakespeare, with an interactive chat for sequence completion.

Character-level fork of Fairseq for sequence-to-sequence learning

Character-level GPT with six-part attention head analysis — ablation, head patching, residual patching, redundancy testing, seed robustness, and logit lens — converging on one load-bearing head.