chargoddard/llama33b-s2a4

LLaMA 33b finetuned on wikitext_document_level with combined linear and NTK-aware ROPE scaling (alpha=4, scale=2.) This model will be coherent up to at least 8k context length, but might work beyond that. This is a merged version of llama33b-s2a4-qlora.

Note that this is not an instruct model - this is base LLaMA with an extended sequence length.