flippa-v2

This model is a fine-tuned version of TheBloke/Mistral-7B-Instruct-v0.2-GPTQ on a mixed dataset of filtered non-refusal data, math, and code. It achieves the following results on the evaluation set:

• Loss: 0.9289

Model description

My second test of experiments using Quantitized LoRA and Mistral-7B-Instruct, trained on A100 in one hour, will increase training times and amount of data as I gain access to more GPUs.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

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
1.5374	0.99	37	1.4226
1.1746	2.0	75	1.2444
1.0746	2.99	112	1.1636
0.9931	4.0	150	1.1037
0.9587	4.99	187	1.0549
0.9101	6.0	225	1.0124
0.8847	6.99	262	0.9782
0.8239	8.0	300	0.9515
0.818	8.99	337	0.9345
0.7882	9.87	370	0.9289

Framework versions

• PEFT 0.9.0
• Transformers 4.38.1
• Pytorch 2.1.0+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2