From f3a9b95912d8b562af96fdedfcf1919744ddf64c Mon Sep 17 00:00:00 2001
From: kazuki <kaz.tokyo.tech20@gmail.com>
Date: Sun, 4 Aug 2024 10:24:21 +0900
Subject: [PATCH] feat: instruct model

---
 .../Llama-3-8B/Llama-3-8B-instruct-v0.2.sh    |  8 ++--
 tools/dataset/convert_dataset_instruct.py     | 42 +++++++++++++++++++
 2 files changed, 46 insertions(+), 4 deletions(-)
 create mode 100644 tools/dataset/convert_dataset_instruct.py

diff --git a/scripts/tsubame/instruct/Llama-3-8B/Llama-3-8B-instruct-v0.2.sh b/scripts/tsubame/instruct/Llama-3-8B/Llama-3-8B-instruct-v0.2.sh
index 30846bf..fa9e3f5 100644
--- a/scripts/tsubame/instruct/Llama-3-8B/Llama-3-8B-instruct-v0.2.sh
+++ b/scripts/tsubame/instruct/Llama-3-8B/Llama-3-8B-instruct-v0.2.sh
@@ -1,7 +1,7 @@
 #!/bin/sh
 #$ -cwd
-#$ -l node_f=2
-#$ -l h_rt=1:00:00:00
+#$ -l node_f=4
+#$ -l h_rt=0:20:00:00
 #$ -o outputs/Llama-3-8b-instruct/$JOB_ID.log
 #$ -e outputs/Llama-3-8b-instruct/$JOB_ID.log
 #$ -p -5
@@ -61,8 +61,8 @@ mkdir -p ${CHECKPOINT_SAVE_DIR}
 # dataset
 DATASET_DIR=/gs/bs/tga-NII-LLM/datasets/raw/instruct/synthetic/general/Synthetic-JP-Conversations-Magpie-Nemotron-4-10k
 
-TRAIN_DATA_PATH=${DATASET_DIR}/Synthetic-JP-Conversations-Magpie-Nemotron-4-10k.jsonl
-VALID_DATA_PATH=${DATASET_DIR}/Synthetic-JP-Conversations-Magpie-Nemotron-4-10k.jsonl
+TRAIN_DATA_PATH=${DATASET_DIR}/converted.jsonl
+VALID_DATA_PATH=${DATASET_DIR}/converted.jsonl
 
 # job name
 JOB_NAME="Llama-3-8B-instruct-v0.2-BS=${GLOBAL_BATCH_SIZE}-LR=${LR}-MINLR=${MIN_LR}-WD=${WEIGHT_DECAY}-GC=${GRAD_CLIP}"
diff --git a/tools/dataset/convert_dataset_instruct.py b/tools/dataset/convert_dataset_instruct.py
new file mode 100644
index 0000000..ed35caf
--- /dev/null
+++ b/tools/dataset/convert_dataset_instruct.py
@@ -0,0 +1,42 @@
+import argparse
+import json
+import copy
+
+
+def convert_jsonl(input_path: str, output_path: str) -> None:
+    converted_data = []
+
+    with open(input_path, 'r', encoding='utf-8') as file:
+        for line in file:
+            item = json.loads(line)
+            messages = item['messages']
+
+            assert len(messages) % 2 == 0
+            conversation_turn: int = len(messages) // 2
+
+            inputs = []
+            for i in range(conversation_turn):
+                user_message = messages[i * 2]
+                assistant_message = messages[i * 2 + 1]
+                inputs.append(user_message)
+                converted_data.append({
+                    "input": copy.deepcopy(inputs),
+                    "output": assistant_message
+                })
+                inputs.append(assistant_message)
+
+    with open(output_path, 'w', encoding='utf-8') as outfile:
+        for item in converted_data:
+            outfile.write(json.dumps(item, ensure_ascii=False) + '\n')
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert JSONL file keys to specified format.")
+    parser.add_argument('--input-path', type=str, help='Path to the input JSONL file')
+    parser.add_argument('--output-path', type=str, help='Path to the output JSONL file')
+
+    args = parser.parse_args()
+
+    convert_jsonl(args.input_path, args.output_path)
+
+if __name__ == "__main__":
+    main()