athirdpath/BigMistral-11b-GLUE_LORA

Regret: Should have not targeted Q, V, K, O; as those are less impactful for "healing" but more impactful on performance otherwise. Still works great!

qlora

This model is a fine-tuned version of athirdpath/BigMistral-11b on the athirdpath/Merge_Glue dataset. It achieves the following results on the evaluation set:

• Loss: 0.9174

Before and After Example

Example model is athirdpath/CleverMage-11b

Example with LoRA (min_p, alpaca)

Example without LoRA (min_p, chatML)

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 10
• eval_batch_size: 10
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 40
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.2198	0.63	30	0.9055
1.1206	1.26	60	0.8951
1.1319	1.89	90	0.8904
1.0031	2.51	120	0.9174

Framework versions

• Transformers 4.35.2
• Pytorch 2.0.1+cu118
• Datasets 2.15.0
• Tokenizers 0.15.0