solar-10b-ocn-v1

This model is a fine-tuned version of upstage/SOLAR-10.7B-v1.0 on the oncc_medqa_instruct dataset.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 4
• eval_batch_size: 8
• 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: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1.0
• mixed_precision_training: Native AMP

Training script

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py --stage sft --do_train True --model_name_or_path upstage/SOLAR-10.7B-v1.0 --template solar --finetuning_type lora --quantization_bit 4 --flash_attn True --dataset_dir data --dataset oncc_medqa_instruct --cutoff_len 1024 --learning_rate 0.0005 --num_train_epochs 1.0 --max_samples 5000 --per_device_train_batch_size 4 --gradient_accumulation_steps 4 --lr_scheduler_type cosine --max_grad_norm 1.0 --logging_steps 10 --save_steps 100 --warmup_steps 10 --neftune_noise_alpha 0.5 --lora_rank 8 --lora_dropout 0.2 --lora_target wqkv --output_dir /workspace/solar-10b-ocn-v1 --fp16 True --plot_loss True

Framework versions

• PEFT 0.8.2
• Transformers 4.37.2
• Pytorch 2.1.1+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1

Performance

Test script: lm_eval --model hf --model_args pretrained=upstage/SOLAR-10.7B-v1.0,peft=chenhugging/solar-10b-ocn-v1,trust_remote_code=True,parallelize=True,load_in_4bit=True --tasks ocn,aocnp,medmcqa,pubmedqa,mmlu_clinical_knowledge,mmlu_college_medicine,mmlu_professional_medicine --device cuda:0 --limit 100

hf (pretrained=upstage/SOLAR-10.7B-v1.0,peft=chenhugging/solar-10b-ocn-v1,trust_remote_code=True,parallelize=True,load_in_4bit=True), gen_kwargs: (None), limit: 100.0, num_fewshot: None, batch_size: 1

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
pubmedqa	1	none	0	acc	0.95	±	0.0219
medmcqa	Yaml	none	0	acc	0.42	±	0.0496
professional_medicine	0	none	0	acc	0.72	±	0.0451
college_medicine	0	none	0	acc	0.67	±	0.0473
clinical_knowledge	0	none	0	acc	0.64	±	0.0482
ocn	Yaml	none	0	acc	0.83	±	0.0378
aocnp	Yaml	none	0	acc	0.72	±	0.0451