mistral-7b-dolphin-sft

mistral-7b-dolphin-sft is an SFT fine-tuned version of unsloth/mistral-7b-bnb-4bit using the cognitivecomputations/dolphin dataset.

Fine-tuning configuration

LoRA

• r: 256
• LoRA alpha: 128
• LoRA dropout: 0.0

Training arguments

• Epochs: 1
• Batch size: 4
• Gradient accumulation steps: 6
• Optimizer: adamw_torch_fused
• Max steps: 100
• Learning rate: 0.0002
• Weight decay: 0.1
• Learning rate scheduler type: linear
• Max seq length: 2048
• 4-bit bnb: True

Trained with Unsloth and Huggingface's TRL library.