harims95/hobbylm-longctx-2p5b

HobbyLM-1B Long-Context Extension Corpus 2.5B GPT-2 tokens staged for extending harims95/hobbylm-1b-hf's context window from 1024 tokens to 4096, then 8192. Format Each <source>_NNNNNN.bin file is the same flat-binary format used throughout HobbyLM training: a 256 x int32 header (magic=20240520, version=1, num_tokens), followed by the token stream as uint16 GPT-2 token ids. The flat .bin files do not encode document boundaries on their own. Each source also ships… See the full description on the dataset page: https://huggingface.co/datasets/harims95/hobbylm-longctx-2p5b.