Artelnics/gpt2-small-opennn

GPT-2 Small for OpenNN

OpenAI's GPT-2 Small converted by Artelnics to OpenNN's flat FP32 parameter format, with a 256-token sequence length. Use it with OpenNN's TextGenerationNetwork and the gpt2 example.

File	Contents
gpt2-small-seq256.bin	OpenNN parameter buffer, 649996672 bytes
vocab.json	Original 50,257-token mapping, serialized for the tokenizer
merges.txt	Original byte-pair merges
LICENSE_GPT2.txt	Complete original OpenAI Modified MIT License
SOURCE.md	Original source, conversion, attribution and SHA-256 digests

Usage

Build OpenNN target gpt2 with CUDA and run it from the executable directory:

./gpt2 "Artificial intelligence"

The example downloads the files into ../data/gpt2, generates at most 40 new tokens and labels the continuation as GPT-2 generated text. Its model dimensions are 256, 50258, 768, 12, 3072, 12; OpenNN reserves [PAD] at index zero.

Original source and licence

Software Copyright (c) 2019 OpenAI. The model retains the original Modified MIT License, including its additional generated-content paragraphs. Preserve LICENSE_GPT2.txt and SOURCE.md with redistributed weights and tokenizer files. OpenNN's software licence does not replace the model licence.

Credit: Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei and Ilya Sutskever, Language Models are Unsupervised Multitask Learners, 2019. See OpenAI's original release and the original model card for training details and limitations. The converted model uses a shorter 256-token context; generated text is not a verified factual source.

File identity

These are the unchanged assets formerly distributed as gpt2-weights-v1. Their sizes and SHA-256 digests are recorded in SOURCE.md. The OpenNN binary uses its compiled parameter layout; match the example's dimensions when loading.