llama3

llama-factory/config/llama3-8b_lora_sft_bf16-p1.yaml

@@ -0,0 +1,47 @@
+### model
+model_name_or_path: FlagAlpha/Llama3-Chinese-8B-Instruct
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+# quantization_bit: 4                     # use 4-bit QLoRA
+loraplus_lr_ratio: 16.0                 # use LoRA+ with lambda=16.0
+# use_unsloth: true                       # use UnslothAI's LoRA optimization for 2x faster training
+upcast_layernorm: true
+
+### dataset
+dataset: alpaca_mgtv_p1
+template: llama3
+cutoff_len: 4096
+max_samples: 25000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/llama3-8b/lora/sft_bf16_p1_full
+logging_steps: 100
+save_steps: 2109
+plot_loss: true
+overwrite_output_dir: true
+# resume_from_checkpoint: true
+
+### train
+per_device_train_batch_size: 16
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 2109
+
+report_to: wandb
+run_name: llama3_8b_p1_full # optional

scripts/tune-mgtv-llama3_8b.sh

@@ -0,0 +1,43 @@
+#!/bin/sh
+
+BASEDIR=$(dirname "$0")
+cd $BASEDIR/..
+echo Current Directory:
+pwd
+
+BASEDIR=`pwd`
+
+nvidia-smi
+uname -a
+cat /etc/os-release
+lscpu
+grep MemTotal /proc/meminfo
+
+#pip install -r requirements.txt
+cd ../LLaMA-Factory && pip install -e .[torch,bitsandbytes] && cd $BASEDIR
+#pip install -U transformers
+#pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
+
+export LOGICAL_REASONING_DATA_PATH=datasets/mgtv
+
+export MODEL_NAME=FlagAlpha/Llama3-Chinese-8B-Instruct
+export MODEL_PREFIX=llama3-8b_lora_sft_bf16
+
+export CONFIG_FILE=config/$MODEL_PREFIX-p1.yaml
+echo "Tuning with $CONFIG_FILE"
+$BASEDIR/scripts/tune-lf.sh $CONFIG_FILE
+
+export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p1.csv
+export ADAPTER_PATH_BASE=llama-factory/saves/qwen2-72b/lora/sft_4bit_p1_full
+echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
+python llm_toolkit/eval_logical_reasoning_all_epochs.py
+
+
+export CONFIG_FILE=config/$MODEL_PREFIX-p2.yaml
+echo "Tuning with $CONFIG_FILE"
+$BASEDIR/scripts/tune-lf.sh $CONFIG_FILE
+
+export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p2.csv
+export ADAPTER_PATH_BASE=llama-factory/saves/qwen2-72b/lora/sft_4bit_p2_full
+echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
+python llm_toolkit/eval_logical_reasoning_all_epochs.py