Merge pull request 'qwen_train' (#2) from qwen_train into main

Reviewed-on: #2

.gitignore

@@ -6,4 +6,6 @@ data
 *.json
 models
 *.log
-research_notebook/data
+research_notebook/data
+train/qwen/output
+train/qwen/mlruns

requirements.txt

@@ -2,4 +2,9 @@ python-dotenv==1.1.1
 hazm=0.10.0
 faiss-cpu==1.12.0
 sentence-transformers==5.1.2
-einops==0.8.1
+einops==0.8.1
+ms-swift
+mteb==1.39.7
+transformers
+liger-kernel
+deepspeed

train/qwen/a.sh

@@ -0,0 +1,35 @@
+# pip install flash-attn --no-build-isolation
+
+
+nproc_per_node=1
+
+MLFLOW_TRACKING_URI=http://0.0.0.0:5004 \
+INFONCE_USE_BATCH=False \
+CUDA_VISIBLE_DEVICES=0 \
+NPROC_PER_NODE=$nproc_per_node \
+swift sft \
+    --model $(pwd)/../../data/models/Qwen3-Embedding-0.6B/model \
+    --task_type embedding \
+    --model_type qwen3_emb \
+    --train_type lora \
+    --lora_rank 8 \
+    --lora_alpha 16 \
+    --target_modules all-linear \
+    --dataset my_local_dataset \
+    --custom_register_path $(pwd)/../../data/dataset/my_dataset_register.py \
+    --split_dataset_ratio 0.005 \
+    --eval_strategy steps \
+    --output_dir output \
+    --eval_steps 1000 \
+    --num_train_epochs 1 \
+    --save_steps 1000 \
+    --save_total_limit 10 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 16 \
+    --gradient_accumulation_steps 4 \
+    --learning_rate 2.4e-5 \
+    --loss_type infonce \
+    --label_names labels \
+    --dataloader_drop_last true \
+    --deepspeed zero3 \
+    --report_to mlflow

train/qwen/merge_model.py

@@ -0,0 +1,24 @@
+import json
+import numpy as np
+import os
+from peft import PeftModel
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
+
+def merge(base_model_path, peft_model_path, save_path):
+    base_model = AutoModelForCausalLM.from_pretrained(base_model_path, torch_dtype="bfloat16")
+    ft_model = PeftModel.from_pretrained(base_model, peft_model_path)
+    ft_model = ft_model.merge_and_unload()
+    ft_model.save_pretrained(save_path)
+
+def main():
+    file_path = os.path.dirname(__file__)
+
+    base_model_path = file_path + "/../../data/models/Qwen3-Embedding-0.6B/model"
+    peft_model_path = file_path + "/output/v1-20251122-184545/checkpoint-3434"
+    save_path = file_path + "/output/v1-20251122-184545/merged_checkpoint-3434"
+    merge(base_model_path, peft_model_path, save_path)
+
+
+if __name__ == "__main__":
+    main()