src/train_sft/examples/debug_augmentation.sh

#!/usr/bin/env bash
# 最小化的数据增强测试脚本，用于调试collator问题

set -euo pipefail

# 激活conda环境
source /home/v-meiszhang/miniconda3/etc/profile.d/conda.sh
conda activate group

# 参考run_group_layout_sft.sh的配置
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")"/../../.. && pwd)"
LF_SRC="$ROOT_DIR/src/train_sft/src"
export PYTHONPATH="$LF_SRC${PYTHONPATH:+:$PYTHONPATH}"

# 使用现有的全量微调配置
YAML="src/train_sft/examples/train_full/group_layout_qwen2_5_full_sft.yaml"

# 设置数据集目录
export FT_DATASET_DIR="${FT_DATASET_DIR:-$ROOT_DIR/datas}"

# 生成时间戳用于输出目录
RUN_ID="debug_aug_$(date +%Y%m%d_%H%M%S)"

# 极简配置，仅测试collator
EXTRA_ARGS=(
  output_dir="$ROOT_DIR/saves/debug_augmentation_$RUN_ID"
  dataset_dir="$ROOT_DIR/src/train_sft/data"
  max_samples=1
  num_train_epochs=1
  per_device_train_batch_size=1
  gradient_accumulation_steps=1
  logging_steps=1
  save_steps=1000
  eval_steps=1000
  warmup_steps=0
  max_steps=1
  enable_data_augmentation=true
  augmentation_ratio=1.0
  augmentation_types="random_group_shuffle"
  overwrite_output_dir=true
  report_to="none"
  evaluation_strategy="no"
  save_strategy="no"
  lr_scheduler_type="constant"
  learning_rate=1e-5
)

# 使用简单的命令格式，避免复杂的torchrun
CMD=(python -m llamafactory.cli train "$YAML" "${EXTRA_ARGS[@]}")

echo "🔍 开始数据增强调试测试"
echo "配置文件: $YAML"
echo "输出目录: $ROOT_DIR/saves/debug_augmentation_$RUN_ID"
echo "执行命令: ${CMD[*]}"
echo ""

# 设置环境变量以减少输出噪音
export FORCE_TORCHRUN=0

# 执行命令
"${CMD[@]}"