4 hard negative trained retriever

thu-coai · Aug 13, 2023 · 7c84016 · 7c84016
1 parent 7020e63
commit 7c84016
Show file tree

Hide file tree

Showing 8 changed files with 8 additions and 8 deletions.
diff --git a/README.md b/README.md
@@ -150,7 +150,7 @@ which will merge the two pairs into `train_lm_0.bin` and `train_lm_0.idx`.
     ```bash
     python3 tools/process_retriever_train_data.py --save retriever_data --data-names TRAIN
     ```
-+ Train the retriever. The `train.jsonl` and `valid.jsonl` data should be put in `retriever_data/TRAIN/p1_en1_hn1_s42/merge`. The trained retriever can be downloaded from this [link](https://huggingface.co/t1101675/PICL/tree/main/results/retriever).
++ Train the retriever. The `train.jsonl` and `valid.jsonl` data should be put in `retriever_data/TRAIN/p1_en1_hn4_s42/merge`. The trained retriever can be downloaded from this [link](https://huggingface.co/t1101675/PICL/tree/main/results/retriever).
     ```bash
     bash scripts/retriever/train.sh ${BASE_PATH}
     ```

diff --git a/scripts/filter/filter.sh b/scripts/filter/filter.sh
@@ -18,7 +18,7 @@ CKPT_NAME="gpt2-large"
 CKPT="${BASE_PATH}/results/${CKPT_NAME}/"
 # data
 RAW_DATA="100K_128"
-SEARCH_DATA="${RAW_DATA}/TRAIN_p1_en1_hn1_s42_lr5e-05-bs64-G1_4375.pt/L2"
+SEARCH_DATA="${RAW_DATA}/TRAIN_p1_en1_hn4_s42_lr5e-05-bs64-G1_4212.pt/L2"
 DATA_DIR="${BASE_PATH}/pretrain_data/${RAW_DATA}/gpt2"
 IDX_DATA_DIR="${BASE_PATH}/pretrain_data/retrieval_results/${SEARCH_DATA}"
 # hp

diff --git a/scripts/pretrain/pretrain_picl_gpt2_large.sh b/scripts/pretrain/pretrain_picl_gpt2_large.sh
@@ -18,7 +18,7 @@ CKPT_NAME="gpt2-large"
 CKPT="${BASE_PATH}/results/${CKPT_NAME}/"
 # data
 CORPUS="100K_128"
-DATA_PREFIX="picl/${CORPUS}_TRAIN_p1_en1_hn1_s42_lr5e-05-bs64-G1_4375.pt_L2_filtered_0.0"
+DATA_PREFIX="picl/${CORPUS}_TRAIN_p1_en1_hn4_s42_lr5e-05-bs64-G1_4212.pt_L2_filtered_0.0"
 LM_DATA_PREFIX="full_doc/"
 DATA_DIR="${BASE_PATH}/pretrain_data/${CORPUS}/gpt2"
 IDX_DATA_DIR="${BASE_PATH}/pretrain_data/${DATA_PREFIX}"

diff --git a/scripts/pretrain/pretrain_picl_gpt2_xlarge.sh b/scripts/pretrain/pretrain_picl_gpt2_xlarge.sh
@@ -18,7 +18,7 @@ CKPT_NAME="gpt2-xlarge"
 CKPT="${BASE_PATH}/results/${CKPT_NAME}/"
 # data
 CORPUS="100K_128"
-DATA_PREFIX="picl/${CORPUS}_TRAIN_p1_en1_hn1_s42_lr5e-05-bs64-G1_4375.pt_L2_filtered_0.0"
+DATA_PREFIX="picl/${CORPUS}_TRAIN_p1_en1_hn4_s42_lr5e-05-bs64-G1_4212.pt_L2_filtered_0.0"
 LM_DATA_PREFIX="full_doc/"
 DATA_DIR="${BASE_PATH}/pretrain_data/${CORPUS}/gpt2"
 IDX_DATA_DIR="${BASE_PATH}/pretrain_data/${DATA_PREFIX}"

diff --git a/scripts/pretrain/pretrain_picl_gpt_neo.sh b/scripts/pretrain/pretrain_picl_gpt_neo.sh
@@ -18,7 +18,7 @@ CKPT_NAME="gpt-neo"
 CKPT="${BASE_PATH}/checkpoints/${CKPT_NAME}/"
 # data
 CORPUS="100K_128"
-DATA_PREFIX="picl/${CORPUS}_TRAIN_p1_en1_hn1_s42_lr5e-05-bs64-G1_4375.pt_L2_filtered_0.0"
+DATA_PREFIX="picl/${CORPUS}_TRAIN_p1_en1_hn4_s42_lr5e-05-bs64-G1_4375.pt_L2_filtered_0.0"
 LM_DATA_PREFIX="full_doc/"
 DATA_DIR="${BASE_PATH}/pretrain_data/${CORPUS}/gpt-j"
 IDX_DATA_DIR="${BASE_PATH}/pretrain_data/${DATA_PREFIX}"

diff --git a/scripts/retriever/infer.sh b/scripts/retriever/infer.sh
@@ -16,7 +16,7 @@ BATCH_SIZE=128
 SEED=10
 MAX_LEN=256
 # runtime
-CKPT=${4-"TRAIN_p1_en1_hn1_s42/lr5e-05-bs64-G1/4375.pt"}
+CKPT=${4-"TRAIN_p1_en1_hn4_s42/lr5e-05-bs64-G1/4212.pt"}
 LOAD_PATH="${WORKING_DIR}/results/retriever/${CKPT}"
 SAVE_PATH="${WORKING_DIR}/pretrain_data/retrieval_results/"
 

diff --git a/scripts/retriever/search.sh b/scripts/retriever/search.sh
@@ -4,7 +4,7 @@ WORKING_DIR=${1}
 
 MODEL_DIR="${WORKING_DIR}/checkpoints/roberta-base/"
 METRIC_TYPE="L2"
-DATA_NAME=${2-"100K_128/TRAIN_p1_en1_hn1_s42_lr5e-05-bs64-G1_4375.pt"}
+DATA_NAME=${2-"100K_128/TRAIN_p1_en1_hn4_s42_lr5e-05-bs64-G1_4212.pt"}
 EMBED_DIR="${WORKING_DIR}/pretrain_data/retrieval_results/${DATA_NAME}"
 SAVE_DIR="${WORKING_DIR}/pretrain_data/retrieval_results/"
 MAX_NUM=-1

diff --git a/scripts/tools/split_picl_train_valid.sh b/scripts/tools/split_picl_train_valid.sh
@@ -1,5 +1,5 @@
 BASE_PATH=${1}
-DATA_PREFIX=${2-"100K_128_TRAIN_p1_en1_hn1_s42_lr5e-05-bs64-G1_4375.pt_L2/filtered_0.0"}
+DATA_PREFIX=${2-"100K_128_TRAIN_p1_en1_hn4_s42_lr5e-05-bs64-G1_4212.pt_L2/filtered_0.0"}
 DATA_NAME="filtered"
 
 MAX_LENGTH=1024