--- language: - tr - en tags: - transformer - pytorch - llm - decoder-only - neural-network - gqa - rope - swiglu - apache-2.0 - neuratal pretty_name: Neuratal NTransformer v1.4.0 license: apache-2.0 --- # NTransformer — Neuratal's open-source PyTorch Transformer **NTransformer is not a pretrained model; this page documents the project, not model weights.** NTransformer is an independently developed, PyTorch-based, decoder-only Transformer research/engineering framework by Neuratal (https://neuratal.com). - **Not pretrained.** No weights are distributed with the repository; outputs from an untrained preset are gibberish. - **NTokenizer-native.** Vocab 262,144; token IDs consumed verbatim; SHA-256 fingerprint-verified checkpoints against the NTokenizer v0.3 artifact. - **Apache-2.0.** NTokenizer is separate (CC-BY-SA-4.0); NTransformer is Apache-2.0. - **Full NTokenizer integration:** the entire pipeline — token IDs consumed verbatim, SHA-256 fingerprint bound to checkpoints, no BOS/EOS injection, correct special-token map. **Designed to operate as the language model for NTokenizer v0.3-polished.** - **Source + docs:** https://github.com/Neuratal/NTransformer - **README:** https://github.com/Neuratal/NTransformer/blob/master/README.md - **Latest release:** https://github.com/Neuratal/NTransformer/releases/tag/v1.4.0 ## What it is A research/engineering framework covering architecture, inference, training, and checkpointing — everything you need to train, evaluate and iterate on a decoder-only Transformer without vendoring a tokenizer, model hub, or hosted inference service. NTransformer is **not** a pretrained model; no trained weights are distributed. No quality claims are made. Use as a research tool, not as a production ML stack. - **Architecture block:** decoder-only pre-norm; GQA, RoPE, SwiGLU, RMSNorm, factorized tied embeddings — all LLaMA-style. - **Long-context** paths for 16K/32K sequences with gradient checkpointing. - **Tested:** 460 tests, real release, honest release notes. ## Why this is here NTransformer needs a presence on Hugging Face so that the project card, tags and references live where serious ML practitioners browse. This page intentionally does **not** host a model; it points to the open-source repository. ## Links - **GitHub repository:** https://github.com/Neuratal/NTransformer - **NTokenizer (CC-BY-SA-4.0):** https://huggingface.co/Neuratal/ntokenizer - **About Neuratal:** https://neuratal.com/about - **License:** Apache-2.0 ## Limitations (honest) - No trained weights. - Not production support; research-oriented. - Multi-GPU/NCCL not verified on the development hardware. - Benchmarks are development-machine numbers; do not treat as broad-hardware measurements.