Nekochu's picture
Update README.md
5915c09 verified
|
raw
history blame
3.68 kB
metadata
model_creator: Nekochu
quantized_by: Nekochu
model_name: Llama-2 13B German ORPO
pretty_name: Llama-2 13B German ORPO
model_type: llama2
prompt_template: >-
  Below is an instruction that describes a task. Write a response that
  appropriately completes the request. ### Instruction: {Instruction} {summary}
  ### input: {category} ### Response: {prompt}
base_model: meta-llama/Llama-2-13b-chat-hf
library_name: peft
license: apache-2.0
datasets:
  - mayflowergmbh/intel_orca_dpo_pairs_de
  - LeoLM/OpenSchnabeltier
  - LeoLM/German_Songs
  - LeoLM/German_Poems
  - bjoernp/ultrachat_de
  - mayflowergmbh/ultra-chat_de
  - mayflowergmbh/airoboros-3.0_de
  - mayflowergmbh/booksum_de
  - mayflowergmbh/dolphin_de
  - mayflowergmbh/evol-instruct_de
  - mayflowergmbh/openschnabeltier_de
  - mayflowergmbh/alpaca-gpt4_de
  - mayflowergmbh/dolly-15k_de
  - mayflowergmbh/oasst_de
language:
  - de
  - en
pipeline_tag: text-generation
task_categories:
  - question-answering
  - text2text-generation
  - conversational
inference: true
tags:
  - llama-factory
  - lora
  - generated_from_trainer
  - llama2
  - llama
  - instruct
  - finetune
  - llm
  - pytorch
  - llama
  - llama-2
  - german
  - deutsch
model-index:
  - name: Llama-2-13B-German-ORPO
    results: []

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