nakcnx/typhoon-sql-qlora

Model Description

Typhoon-7b QLoRA Finetune by unsloth with SQL Context dataset.

Training Hyperparameters

Batch Size: 48 (4(BS)x4(GAS)x3(GPU))
The following bitsandbytes quantization config was used during training:

• r = 64
• lora_alpha = 16
• lora_dropout = 0.05
• 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: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: float16

LOSS

Step Training Loss Eval Loss
1550 (Epoch:1) 0.4295 0.4367
3110 (Epoch:2) 0.4057 0.4217

Framework versions

• PEFT 0.7.0