This represents the PEFT weights only. The base model is LLaMA 2. Instruction finetuning was done using 4 bit QLoRA on a single A100 GPU with the PEFT config as given below. The dataset used for this instruction finetuning process is a translated version of the cleaned alpaca dataset (translated using NLLB-1.3B).

Do note that this model might have inferior performance on some language specific tasks compared to full finetuning or a different base model trained with more language specific data.

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