Add re-sft scripts

eval-scripts/eval_13b_base.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+
+set -eux
+
+SCRIPT_PATH="llm-jp-eval/"
+cd $SCRIPT_PATH
+
+source venv/bin/activate
+export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
+
+# Evaluate procedure
+function evaluate_13b_model () {
+    TARGET_MODEL=$1
+    WANDB_NAME=$2
+
+    WANDB_ENTITY="llm-jp"
+    WANDB_PROJECT="13b-sft-and-eval"
+
+    python scripts/evaluate_llm.py \
+        model.pretrained_model_name_or_path="$TARGET_MODEL" \
+        tokenizer.pretrained_model_name_or_path="$TARGET_MODEL" \
+        target_dataset=all \
+        wandb.run_name="$WANDB_NAME" \
+        wandb.entity="$WANDB_ENTITY" \
+        wandb.project="$WANDB_PROJECT" \
+        wandb.log=true
+}
+
+# exp6
+TARGET_MODEL="/model/13B_HF/llm-jp-13b-cc-v2-63500step.code10K_en20K_ja30K_ver2.2" # exp6
+WANDB_NAME="eval-exp6-base"
+evaluate_13b_model $TARGET_MODEL $WANDB_NAME
+
+# exp8
+TARGET_MODEL="/model/13B_HF/llm-jp-13b-cc-v2-beta-61000step.code20K_en40K_ja60K_ver2.2" # exp8
+WANDB_NAME="eval-exp8-base"
+evaluate_13b_model $TARGET_MODEL $WANDB_NAME

eval-scripts/eval_13b_sft-full-jaster.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+
+set -eux
+
+SCRIPT_PATH="llm-jp-eval/"
+cd $SCRIPT_PATH
+
+source venv/bin/activate
+export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
+
+# Evaluate procedure
+function evaluate_13b_model () {
+    TARGET_MODEL=$1
+    WANDB_NAME=$2
+
+    WANDB_ENTITY="llm-jp"
+    WANDB_PROJECT="13b-sft-and-eval"
+
+    python scripts/evaluate_llm.py \
+        model.pretrained_model_name_or_path="$TARGET_MODEL" \
+        tokenizer.pretrained_model_name_or_path="$TARGET_MODEL" \
+        target_dataset=all \
+        wandb.run_name="$WANDB_NAME" \
+        wandb.entity="$WANDB_ENTITY" \
+        wandb.project="$WANDB_PROJECT" \
+        wandb.log=true
+}
+
+# exp6
+TARGET_MODEL="/model/13B_HF_RE/llm-jp-13b-cc-v2-63500step.code10K_en20K_ja30K_ver2.2_full-jaster" # exp6
+WANDB_NAME="eval-exp6-full-jaster"
+evaluate_13b_model $TARGET_MODEL $WANDB_NAME
+
+# exp8
+TARGET_MODEL="/model/13B_HF_RE/llm-jp-13b-cc-v2-beta-61000step.code20K_en40K_ja60K_ver2.2_full-jaster" # exp8
+WANDB_NAME="eval-exp8-full-jaster"
+evaluate_13b_model $TARGET_MODEL $WANDB_NAME

eval-scripts/eval_7b_base.sh

@@ -0,0 +1,66 @@
+#!/bin/bash
+
+set -eux
+
+# SCRIPT_PATH="llm-jp-eval/"
+SCRIPT_PATH="/model/sosuke/llm-jp-eval-v1.1.0"
+cd $SCRIPT_PATH
+
+source venv/bin/activate
+export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
+
+# Evaluate procedure
+function evaluate_7b_model () {
+    TARGET_MODEL=$1
+    WANDB_NAME=$2
+
+    WANDB_ENTITY="llm-jp"
+    WANDB_PROJECT="7b-sft-and-eval"
+
+    python scripts/evaluate_llm.py \
+        model.pretrained_model_name_or_path="$TARGET_MODEL" \
+        tokenizer.pretrained_model_name_or_path="$TARGET_MODEL" \
+        target_dataset=all \
+        wandb.run_name="$WANDB_NAME" \
+        wandb.entity="$WANDB_ENTITY" \
+        wandb.project="$WANDB_PROJECT" \
+        wandb.log=true
+}
+
+# # exp1
+# TARGET_MODEL="/model/7B_HF/llm-jp-7b-61000step.code20K_en40K_ja60K_ver2.2" # exp1
+# WANDB_NAME="eval-exp1-base"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp2
+# TARGET_MODEL="/model/7B_HF/llm-jp-7b-okazaki-lab-cc-63500step.code10K_en20K_ja30K_ver2.2" # exp2
+# WANDB_NAME="eval-exp2-base"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp3
+# TARGET_MODEL="/model/7B_HF/llm-jp-7b-63500step.code10K_en20K_ja30K_ver2.2" # exp3
+# WANDB_NAME="eval-exp3-base"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp4
+# TARGET_MODEL="/model/7B_HF/llm-jp-7b-cc-v2-63500step.code10K_en20K_ja30K_ver2.2" # exp4
+# WANDB_NAME="eval-exp4-base"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp4+
+# TARGET_MODEL="/model/7B_HF/llm-jp-7b-cc-v2-beta-63500step.code10K_en20K_ja30K_ver2.2" # exp4+
+# WANDB_NAME="eval-exp4+-base"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp5
+# # Skip for now
+
+# # exp7
+# TARGET_MODEL="/model/7B_HF/llm-jp-7b-cc-v2-beta-119000step.code10k_en20k_ja30k_ver2.2" # exp7
+# WANDB_NAME="eval-exp7-base"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# exp9
+TARGET_MODEL="/model/7B_HF/llm-jp-7b-CC_v2_100k_ver3.0" # exp9
+WANDB_NAME="eval-exp9-base"
+evaluate_7b_model $TARGET_MODEL $WANDB_NAME

eval-scripts/eval_7b_sft-full-jaster.sh

@@ -0,0 +1,66 @@
+#!/bin/bash
+
+set -eux
+
+# SCRIPT_PATH="llm-jp-eval/"
+SCRIPT_PATH="/model/sosuke/llm-jp-eval-v1.1.0"
+cd $SCRIPT_PATH
+
+source venv/bin/activate
+export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
+
+# Evaluate procedure
+function evaluate_7b_model () {
+    TARGET_MODEL=$1
+    WANDB_NAME=$2
+
+    WANDB_ENTITY="llm-jp"
+    WANDB_PROJECT="7b-sft-and-eval"
+
+    python scripts/evaluate_llm.py \
+        model.pretrained_model_name_or_path="$TARGET_MODEL" \
+        tokenizer.pretrained_model_name_or_path="$TARGET_MODEL" \
+        target_dataset=all \
+        wandb.run_name="$WANDB_NAME" \
+        wandb.entity="$WANDB_ENTITY" \
+        wandb.project="$WANDB_PROJECT" \
+        wandb.log=true
+}
+
+# # exp1
+# TARGET_MODEL="/model/7B_HF_RE/llm-jp-7b-61000step.code20K_en40K_ja60K_ver2.2_full-jaster" # exp1
+# WANDB_NAME="eval-exp1-full-jaster"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp2
+# TARGET_MODEL="/model/7B_HF_RE/llm-jp-7b-okazaki-lab-cc-63500step.code10K_en20K_ja30K_ver2.2_full-jaster" # exp2
+# WANDB_NAME="eval-exp2-full-jaster"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp3
+# TARGET_MODEL="/model/7B_HF_RE/llm-jp-7b-63500step.code10K_en20K_ja30K_ver2.2_full-jaster" # exp3
+# WANDB_NAME="eval-exp3-full-jaster"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp4
+# TARGET_MODEL="/model/7B_HF_RE/llm-jp-7b-cc-v2-63500step.code10K_en20K_ja30K_ver2.2_full-jaster" # exp4
+# WANDB_NAME="eval-exp4-full-jaster"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp4+
+# TARGET_MODEL="/model/7B_HF_RE/llm-jp-7b-cc-v2-beta-63500step.code10K_en20K_ja30K_ver2.2_full-jaster" # exp4+
+# WANDB_NAME="eval-exp4+-full-jaster"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# # exp5
+# # Skip for now
+
+# # exp7
+# TARGET_MODEL="/model/7B_HF_RE/llm-jp-7b-cc-v2-beta-119000step.code10k_en20k_ja30k_ver2.2_full-jaster" # exp7
+# WANDB_NAME="eval-exp7-full-jaster"
+# evaluate_7b_model $TARGET_MODEL $WANDB_NAME
+
+# exp9
+TARGET_MODEL="/model/7B_HF_RE/llm-jp-7b-CC_v2_100k_ver3.0_full-jaster" # exp9
+WANDB_NAME="eval-exp9-full-jaster"
+evaluate_7b_model $TARGET_MODEL $WANDB_NAME

sft-scripts/README.md

@@ -0,0 +1,67 @@
+# SFT scripts
+
+## 実行環境
+
+MDXの`llm-jp-nvlink`環境で実行
+
+## Full parameter SFT
+
+Full parameter SFTは以下の設定に固定して行う
+
+```yaml
+max_seq_length: 4096   # llama model
+global_batch_size: 64  # = per_device_batch_size * num_devices * gradient_accumulation_steps
+num_epochs: 5
+learning_rate: 2e-5
+lr_scheduler: cosine
+warmup_ratio: 0.1
+dtype: bfloat16
+```
+
+datasetは以下の2種類についてそれぞれ行う
+
+- `jaster`: llm-jp-evalで利用するモデル向けのデータセット
+- `gpt4-self-inst`: VicunaQAで利用するモデル向けのデータセット
+
+### 7B models
+
+上記の設定に加え、以下の設定で学習
+
+- single node (8 GPUs => `gradient_accumulation_steps` = 8)
+- ZeRO-2
+
+### 13B models
+
+上記の設定に加え、以下の設定で学習
+
+- single node (8 GPU => `gradient_accumulation_steps` = 8)
+- ZeRO-3
+- gradient_checkpointingを有効化
+
+## LoRA SFT
+
+LoRA SFTは以下の設定に固定して行う（基本的な設定はFull parameter SFTと共通、LRのみ2種類用意）
+
+```yaml
+max_seq_length: 4096   # llama model
+global_batch_size: 64  # = per_device_batch_size * num_devices * gradient_accumulation_steps
+num_epochs: 5
+learning_rate: [2e-5, 1e-4]
+lr_scheduler: cosine
+warmup_ratio: 0.1
+dtype: bfloat16
+```
+
+datasetは以下の2種類についてそれぞれ行う（Full parameter SFTと共通）
+
+- `jaster`: llm-jp-evalで利用するモデル向けのデータセット
+- `gpt4-self-inst`: VicunaQAで利用するモデル向けのデータセット
+
+### 7B models
+
+すべてのモデルについてsingle GPUで学習
+
+### 13B models
+
+`jaster`のモデルについてはsingle GPUで学習
+`gpt4-self-inst`のモデルについてはメモリ不足によりsingle GPUでは学習ができなかったため、single node (8 GPU) で学習

sft-scripts/exp1_gpt4-self-inst_full.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+
+# Single node full parameter sft script
+
+set -eux
+
+# Move to script path
+SCRIPT_PATH="llm-jp-sft/"
+cd $SCRIPT_PATH
+source venv/bin/activate
+
+# Target Model (TODO: Change if needed)
+TARGET_MODEL="/model/7B_HF/llm-jp-7b-61000step.code20K_en40K_ja60K_ver2.2" # exp1
+SFT_TYPE="full-gpt4-self-inst"
+OUTPUT_DIR="/model/7B_HF_RE/$(basename ${TARGET_MODEL%/})_${SFT_TYPE}"
+
+# wandb info
+export WANDB_ENTITY="llm-jp"
+export WANDB_PROJECT="7b-sft-and-eval"
+export WANDB_NAME="sft-exp1-${SFT_TYPE}"
+
+# For debugging
+export NCCL_DEBUG=WARN
+
+# Training settings
+config_file="configs/accelerate_config_zero3.yaml"
+dataset_path="./dataset"
+dataset_sh="./mdx/dataset_gpt4_self_inst_ja.sh"
+num_train_epochs=5
+per_device_train_batch_size=1
+gradient_accumulation_steps=8  # global_batch_size = per_device_train_batch_size * n_gpus * gradient_accumulation_steps = 64
+max_seq_length=4096             # for llama model
+learning_rate="2e-5"
+lr_scheduler_type="cosine"
+warmup_ratio=0.1
+
+# N_gpu = 8
+accelerate launch --config_file $config_file \
+    train.py \
+    --model_name_or_path $TARGET_MODEL \
+    --tokenizer_name_or_path $TARGET_MODEL \
+    --num_train_epochs $num_train_epochs \
+    --per_device_train_batch_size $per_device_train_batch_size \
+    --gradient_accumulation_steps $gradient_accumulation_steps \
+    --learning_rate $learning_rate \
+    --warmup_ratio $warmup_ratio \
+    --lr_scheduler_type $lr_scheduler_type \
+    --bf16 \
+    --max_seq_length $max_seq_length \
+    --logging_steps 10 \
+    --report_to wandb \
+    --data_files $($dataset_sh $dataset_path) \
+    --output_dir "$OUTPUT_DIR"

sft-scripts/exp1_gpt4-self-inst_lora_all.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+
+# Single GPU sft script
+# based on https://github.com/llm-jp/llm-jp-sft/blob/main/mdx/train_peft_single_gpu.sh
+
+set -eux
+
+# Move to script path
+SCRIPT_PATH="llm-jp-sft/"
+cd $SCRIPT_PATH
+source venv/bin/activate
+
+# Set CUDA Device (default: 0)
+export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0}"
+
+# Target Model (TODO: Change if needed)
+TARGET_MODEL="/model/7B_HF/llm-jp-7b-61000step.code20K_en40K_ja60K_ver2.2" # exp1
+SFT_TYPE="lora-all-gpt4-self-inst"
+OUTPUT_DIR="/model/7B_HF_RE/$(basename ${TARGET_MODEL%/})_${SFT_TYPE}"
+
+# wandb info
+export WANDB_ENTITY="llm-jp"
+export WANDB_PROJECT="7b-sft-and-eval"
+export WANDB_NAME="sft-exp1-${SFT_TYPE}"
+
+# Training settings
+dataset_path="./dataset"
+dataset_sh="./mdx/dataset_gpt4_self_inst_ja.sh"
+num_train_epochs=5
+per_device_train_batch_size=1
+gradient_accumulation_steps=64  # global_batch_size = per_device_train_batch_size * gradient_accumulation_steps = 64
+peft_target_model="llama-all"
+max_seq_length=4096             # for llama model
+learning_rate="2e-5"
+lr_scheduler_type="cosine"
+warmup_ratio=0.1
+
+python train.py \
+    --model_name_or_path $TARGET_MODEL \
+    --tokenizer_name_or_path $TARGET_MODEL \
+    --use_peft \
+    --num_train_epochs $num_train_epochs \
+    --per_device_train_batch_size $per_device_train_batch_size \
+    --gradient_accumulation_steps $gradient_accumulation_steps \
+    --learning_rate $learning_rate \
+    --warmup_ratio $warmup_ratio \
+    --lr_scheduler_type $lr_scheduler_type \
+    --bf16 \
+    --max_seq_length $max_seq_length \
+    --logging_steps 10 \
+    --report_to wandb \
+    --data_files $($dataset_sh $dataset_path) \
+    --output_dir "$OUTPUT_DIR" \
+    --peft_target_model $peft_target_model