chargoddard/llama33b-s2a4-qlora

LLaMA 33b finetuned on wikitext_document_level with a combination of both linear and NTK-aware ROPE scaling.

Trained with alpha=4, scale=2. Definitely works for sequence lengths up to and including 4096. Might work for much longer, but I don't have the VRAM to test properly. ¯\_(ツ)_/¯

Training procedure

The following bitsandbytes quantization config was used during training:

• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: bfloat16

Framework versions

• PEFT 0.4.0.dev0