DocGPT-ft

This model is a fine-tuned version of TheBloke/Mistral-7B-Instruct-v0.2-GPTQ on the lavita/ChatDoctor-HealthCareMagic-100k dataset.

Model description

Uses parameter efficient fine-tuning for QLora

Intended uses & limitations

The intended use is just for fun.

Training and evaluation data

The training set was 90% of the data and testing set was 10%. Only a small percentage of the data was used to reduce training time.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
2.4174	0.9412	12	2.2924
2.1327	1.9608	25	2.2750
2.0864	2.9804	38	2.2745
2.0362	4.0	51	2.2761
2.1357	4.9412	63	2.2849
1.942	5.9608	76	2.2961
1.8904	6.9804	89	2.3165
1.8585	8.0	102	2.3295
1.9923	8.9412	114	2.3390
1.6331	9.4118	120	2.3387

Framework versions

• PEFT 0.12.0
• Transformers 4.42.4
• Pytorch 2.4.0+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1