chore: import upstream snapshot with attribution

2026-07-13 13:24:13 +08:00
commit 1037506f2e
6050 changed files with 1731598 additions and 0 deletions
@@ -0,0 +1,55 @@
+#!/usr/bin/env bash
+
+set -x
+set -e
+
+DIR="$( cd "$( dirname "$0" )" && cd ../../ && pwd )"
+echo "working directory: ${DIR}"
+
+if [ -z "$OUTPUT_DIR" ]; then
+  OUTPUT_DIR="${DIR}/checkpoint/nq_$(date +%F-%H%M.%S)"
+fi
+if [ -z "$DATA_DIR" ]; then
+  DATA_DIR="${DIR}/data/dpr/"
+fi
+
+mkdir -p "${OUTPUT_DIR}"
+
+PROC_PER_NODE=$(nvidia-smi --list-gpus | wc -l)
+# python -u -m torch.distributed.launch --nproc_per_node ${PROC_PER_NODE} src/train_biencoder.py \
+deepspeed src/train_biencoder.py --deepspeed dpr_ds_config.json \
+    --model_name_or_path intfloat/simlm-base-wiki100w \
+    --task_type qa \
+    --per_device_train_batch_size 8 \
+    --per_device_eval_batch_size 32 \
+    --l2_normalize True \
+    --t 0.02 \
+    --t_warmup True \
+    --seed 5678 \
+    --do_train \
+    --fp16 \
+    --train_file "${DATA_DIR}/nq_train.jsonl,${DATA_DIR}/nq_hard_train.jsonl" \
+    --validation_file "${DATA_DIR}/nq_dev.jsonl" \
+    --q_max_len 32 \
+    --p_max_len 192 \
+    --train_n_passages 16 \
+    --use_first_positive True \
+    --dataloader_num_workers 1 \
+    --learning_rate 1e-5 \
+    --use_scaled_loss True \
+    --loss_scale 1 \
+    --max_steps 20000 \
+    --warmup_steps 1000 \
+    --share_encoder True \
+    --logging_steps 50 \
+    --output_dir "${OUTPUT_DIR}" \
+    --data_dir "${DATA_DIR}" \
+    --save_total_limit 3 \
+    --save_strategy steps \
+    --save_steps 2000 \
+    --evaluation_strategy steps \
+    --eval_steps 2000 \
+    --remove_unused_columns False \
+    --overwrite_output_dir \
+    --disable_tqdm True \
+    --report_to none "$@"