README.md

metadata

license: apache-2.0
base_model: NousResearch/Meta-Llama-3.1-8B-Instruct
library_name: peft
tags:
  - llama-factory
  - lora
datasets:
  - Nekochu/novel17_train_alpaca_format
  - bofenghuang/vigogne
  - jpacifico/French-Alpaca-dataset-Instruct-110K
  - MaziyarPanahi/french_instruct_human_sharegpt
  - Snit/french-conversation
language:
  - fr
  - en

• Similar to the old Nekochu/Llama-2-13B-fp16-french with additional datasets.
• I've (alway) kept LoRA QLoRA_french_sft so it can be applied to any LLaMA-3.1-8B fine-tuned model but may affect performance.

This training can be replicated using LLaMA-Factory.

Stage A: Pre Training, Raw text

set CUDA_VISIBLE_DEVICES=0 && llamafactory-cli train --stage pt --do_train True --model_name_or_path NousResearch/Meta-Llama-3.1-8B-Instruct --preprocessing_num_workers 16 --finetuning_type lora --template alpaca --rope_scaling linear --flash_attn fa2 --dataset_dir data --dataset french-raw-pt --cutoff_len 8192 --learning_rate 5e-05 --num_train_epochs 3.0 --max_samples 10000000 --per_device_train_batch_size 1 --gradient_accumulation_steps 1 --lr_scheduler_type cosine --max_grad_norm 1.0 --logging_steps 10 --save_steps 1000 --warmup_steps 0 --neftune_noise_alpha 5 --optim adamw_8bit --packing True --report_to none --output_dir saves\LLaMA3.1-8B-Chat\lora\QLoRA_french_pt --bf16 True --plot_loss True --ddp_timeout 180000000 --include_num_input_tokens_seen True --quantization_bit 4 --quantization_method bitsandbytes --lora_rank 32 --lora_alpha 64 --lora_dropout 0.15 --create_new_adapter True --lora_target all

Stage B: Continued Supervised Fine-Tuning, QA

set CUDA_VISIBLE_DEVICES=0 && llamafactory-cli train --stage sft --do_train True --model_name_or_path NousResearch/Meta-Llama-3.1-8B-Instruct --preprocessing_num_workers 16 --finetuning_type lora --template alpaca --rope_scaling linear --flash_attn fa2 --dataset_dir data --dataset Acquiesce_french_vigogne,novel17_train --cutoff_len 8192 --learning_rate 5e-05 --num_train_epochs 3.0 --max_samples 10000000 --per_device_train_batch_size 1 --gradient_accumulation_steps 1 --lr_scheduler_type cosine --max_grad_norm 1.0 --logging_steps 10 --save_steps 1000 --warmup_steps 0 --neftune_noise_alpha 5 --optim adamw_8bit --packing True --report_to none --output_dir saves\LLaMA3.1-8B-Chat\lora\QLoRA_french_sft --bf16 True --plot_loss True --ddp_timeout 180000000 --adapter_name_or_path saves\LLaMA3.1-8B-Chat\lora\QLoRA_french_pt --quantization_bit 4 --quantization_method bitsandbytes --lora_rank 32 --lora_alpha 64 --lora_dropout 0.15 --lora_target all

Dataset convert to Alpaca: Acquiesce_french_vigogne,french-raw-pt