Nekochu/Llama-2-13B-German-ORPO

llama2 Chat LoRa sft train Stage A on German dataset:

German_Songs,German_Poems,bjoernp_ultrachat_de,OpenSchnabeltier,ultrachat_de,oasst_de,dolly_15k_de,alpaca-gpt4_de,openschnabeltier_de,evol_instruct_de,dolphin_de,booksum_de,airoboros_de & eval VAGOsolutions/MT-Bench-TrueGerman?

Stage B: Resume LoRa training using ORPO and dataset mayflowergmbh/intel_orca_dpo_pairs_de

Oh and I am not GER speaker ^^

Training hyperparameters

python src/train_bash.py --stage sft ... --finetuning_type lora --quantization_bit 4 --template alpaca --rope_scaling linear --flash_attn True --dataset_dir data --dataset German_Songs,German_Poems,bjoernp_ultrachat_de,OpenSchnabeltier,ultrachat_de,oasst_de,dolly_15k_de,alpaca-gpt4_de,openschnabeltier_de,evol_instruct_de,dolphin_de,booksum_de,airoboros_de --cutoff_len 4096 --learning_rate 5e-05 --num_train_epochs 1.0 --max_samples 100000 --per_device_train_batch_size 1 --gradient_accumulation_steps 1 --lr_scheduler_type cosine --max_grad_norm 1.0 --logging_steps 5 --save_steps 1000 --warmup_steps 0 --neftune_noise_alpha 0.5 --optim adamw_torch --upcast_layernorm True --use_llama_pro True --bf16 True --lora_rank 512 --lora_alpha 1024 --lora_dropout 0.15 --lora_target all --use_rslora True --additional_target all --create_new_adapter True --plot_loss True

python src/train_bash.py --stage orpo ... --finetuning_type lora --quantization_bit 4 --template alpaca --rope_scaling linear --flash_attn True --dataset_dir data --dataset orca_dpo_de --cutoff_len 4096 --learning_rate 1e-05 --num_train_epochs 1.0 --max_samples 100000 --per_device_train_batch_size 1 --gradient_accumulation_steps 1 --lr_scheduler_type cosine --max_grad_norm 0.9 --logging_steps 5 --save_steps 250 --warmup_steps 100 --neftune_noise_alpha 0.5 --optim adamw_torch --upcast_layernorm True --use_llama_pro True --report_to none --bf16 True --lora_rank 512 --lora_alpha 1024 --lora_dropout 0.15 --use_rslora True --lora_target all --additional_target all --orpo_beta 0.1 --plot_loss True

The following hyperparameters were used during training:

• learning_rate: 1e-05 # not Defaut LR as for high rank 512, alpha 1024
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 1000
• num_epochs: 1.0

Framework versions

• PEFT 0.10.0
• Transformers 4.39.1
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2