This model was fine-tuned using 4-bit QLoRa, following the instructions in https://huggingface.co/blog/llama2#fine-tuning-with-peft.

The dataset includes 10k prompts.

I used a Amazon EC2 g5.xlarge instance (1xA10G GPU), with the Deep Learning AMI for PyTorch. Training time was about 10 hours. On-demand price is about $10, which can easily be reduced to about $3 with EC2 Spot Instances.

The full log is included, as well as a simple inference script.

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• 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: fp4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: float32

Framework versions

• PEFT 0.5.0