add convert dataset and train qwen

.gitignore

@@ -1,5 +1,5 @@
 data_preprocess/data/*
-data
+data/models
 */__pycache__/*
 .env
 .venv
@@ -9,3 +9,13 @@ models
 research_notebook/data
 train/qwen/output
 train/qwen/mlruns
+output
+data/dataset/__pycache__
+data/dataset/generated_250000_general/__pycache__
+data/dataset/generated_250000_general/generated_250000_general.jsonl
+data/dataset/generated_250000_religous/250_religous_ready.jsonl
+!data/dataset/generated_250000_religous/convert_to_jsonl.py
+data/dataset/generated_250000_religous/__pycache__
+data/dataset/my_local_dataset/__pycache__
+data/dataset/v11_dataset_hn/__pycache__
+data/dataset/v11_generated/__pycache__

data/dataset/generated_250000_general/convert_to_jsonl.py

@@ -0,0 +1,14 @@
+import json
+import os
+
+file_path = os.path.dirname(__file__)
+input_file = file_path + "/generated_250000_general.json"
+output_file = file_path + "/generated_250000_general.jsonl"
+
+with open(input_file, "r", encoding="utf-8") as f_in, open(output_file, "w", encoding="utf-8") as f_out:
+    data = json.load(f_in)  # لیست رکوردها
+    for record in data:
+        json_line = json.dumps(record, ensure_ascii=False)
+        f_out.write(json_line + "\n")
+
+print(f"Converted {input_file} to {output_file}")

data/dataset/generated_250000_general/generated.py

@@ -0,0 +1,57 @@
+from swift.llm import ResponsePreprocessor, DatasetMeta, register_dataset, SubsetDataset, load_dataset
+from typing import Dict, Any
+import os
+
+
+class CustomPreprocessor(ResponsePreprocessor):
+    # def __init__(self, *, columns = None, **kwargs):
+    #     super().__init__(columns=columns, **kwargs)
+    #     self.num_all_negative = 0
+    def get_detailed_instruct(self, task_description: str, query: str) -> str:
+        return f'Instruct: {task_description}\nQuery:{query}'
+
+    def add_template(self, text):
+        task = 'Given a web search query, retrieve relevant passages that answer the query'
+        return self.get_detailed_instruct(task, text)
+
+    def preprocess(self, row: Dict[str, Any]) -> Dict[str, Any]:
+        query = self.add_template(row["query"])
+        passage_positive = row["passage_positive"]
+        passage_negative = row["passage_negative"]
+        passage_negative_random = row["passage_negative_random"]
+
+        all_neg = passage_negative + passage_negative_random
+        all_neg = list(set(all_neg))
+        # self.num_all_negative += len(all_neg)
+
+        row = {
+            # 'query': [{'role': 'user', 'content': query, 'loss': None}],
+            'query': query,
+            'positive_messages': [
+                 [{'role': 'user', 'content': passage_positive[i]}] for i in range(len(passage_positive))
+                ],
+            'negative_messages': [
+                 [{'role': 'user', 'content': all_neg[i]}] for i in range(len(all_neg))
+                ],
+            # 'label': 1.0
+        }
+        if len(row["negative_messages"]) == 0:
+            del row["negative_messages"]
+        return super().preprocess(row)
+
+
+register_dataset(
+    DatasetMeta(
+        dataset_path=os.path.dirname(__file__) + '/generated_250000_general.jsonl',
+        dataset_name="generated_250000_general",
+        # subsets=[SubsetDataset('train', split=['train']), SubsetDataset('test', split=['test'])],
+        preprocess_func=CustomPreprocessor(),
+    ))
+
+if __name__ == '__main__':
+    # load_dataset returns train_dataset and val_dataset based on `split_dataset_ratio`
+    # Here, since we didn't pass `split_dataset_ratio` (defaults to 0), we take the first one (index 0)
+    dataset = load_dataset('generated_250000_general')[0]
+    test_dataset = load_dataset('swift/financial_classification:test')[0]
+    print(f'dataset[0]: {dataset[0]}')
+    print(f'test_dataset[0]: {test_dataset[0]}')

data/dataset/generated_250000_religous/convert_to_jsonl.py

@@ -0,0 +1,14 @@
+import json
+import os
+
+file_path = os.path.dirname(__file__)
+input_file = file_path + "/250_religous_ready.json"
+output_file = file_path + "/250_religous_ready.jsonl"
+
+with open(input_file, "r", encoding="utf-8") as f_in, open(output_file, "w", encoding="utf-8") as f_out:
+    data = json.load(f_in)  # لیست رکوردها
+    for record in data:
+        json_line = json.dumps(record, ensure_ascii=False)
+        f_out.write(json_line + "\n")
+
+print(f"Converted {input_file} to {output_file}")

data/dataset/generated_250000_religous/generated.py

@@ -0,0 +1,57 @@
+from swift.llm import ResponsePreprocessor, DatasetMeta, register_dataset, SubsetDataset, load_dataset
+from typing import Dict, Any
+import os
+
+
+class CustomPreprocessor(ResponsePreprocessor):
+    # def __init__(self, *, columns = None, **kwargs):
+    #     super().__init__(columns=columns, **kwargs)
+    #     self.num_all_negative = 0
+    def get_detailed_instruct(self, task_description: str, query: str) -> str:
+        return f'Instruct: {task_description}\nQuery:{query}'
+
+    def add_template(self, text):
+        task = 'Given a web search query, retrieve relevant passages that answer the query'
+        return self.get_detailed_instruct(task, text)
+
+    def preprocess(self, row: Dict[str, Any]) -> Dict[str, Any]:
+        query = self.add_template(row["query"])
+        passage_positive = row["passage_positive"]
+        passage_negative = row["passage_negative"]
+        passage_negative_random = row["passage_negative_random"]
+
+        all_neg = passage_negative + passage_negative_random
+        all_neg = list(set(all_neg))
+        # self.num_all_negative += len(all_neg)
+
+        row = {
+            # 'query': [{'role': 'user', 'content': query, 'loss': None}],
+            'query': query,
+            'positive_messages': [
+                 [{'role': 'user', 'content': passage_positive[i]}] for i in range(len(passage_positive))
+                ],
+            'negative_messages': [
+                 [{'role': 'user', 'content': all_neg[i]}] for i in range(len(all_neg))
+                ],
+            # 'label': 1.0
+        }
+        if len(row["negative_messages"]) == 0:
+            del row["negative_messages"]
+        return super().preprocess(row)
+
+
+register_dataset(
+    DatasetMeta(
+        dataset_path=os.path.dirname(__file__) + '/250_religous_ready.jsonl',
+        dataset_name="generated_250000_religous",
+        # subsets=[SubsetDataset('train', split=['train']), SubsetDataset('test', split=['test'])],
+        preprocess_func=CustomPreprocessor(),
+    ))
+
+if __name__ == '__main__':
+    # load_dataset returns train_dataset and val_dataset based on `split_dataset_ratio`
+    # Here, since we didn't pass `split_dataset_ratio` (defaults to 0), we take the first one (index 0)
+    dataset = load_dataset('generated_250000_religous')[0]
+    test_dataset = load_dataset('swift/financial_classification:test')[0]
+    print(f'dataset[0]: {dataset[0]}')
+    print(f'test_dataset[0]: {test_dataset[0]}')

data/dataset/my_local_dataset/my_dataset_register.py

@@ -0,0 +1,58 @@
+from swift.llm import ResponsePreprocessor, DatasetMeta, register_dataset, SubsetDataset, load_dataset
+from typing import Dict, Any
+import os
+
+class CustomPreprocessor(ResponsePreprocessor):
+    def __init__(self, *, columns = None, **kwargs):
+        super().__init__(columns=columns, **kwargs)
+        self.num_all_negative = 0
+
+    def get_detailed_instruct(self, task_description: str, query: str) -> str:
+        return f'Instruct: {task_description}\nQuery:{query}'
+
+    def add_template(self, text):
+        task = 'Given a web search query, retrieve relevant passages that answer the query'
+        return self.get_detailed_instruct(task, text)
+
+    def preprocess(self, row: Dict[str, Any]) -> Dict[str, Any]:
+        query = self.add_template(row["query"])
+        passage_positive = row["passage_positive"]
+        passage_negative = row["passage_negative"]
+        passage_negative_random = row["passage_negative_random"]
+
+        all_neg = passage_negative + passage_negative_random
+        all_neg = list(set(all_neg))
+
+        self.num_all_negative += len(all_neg)
+
+        row = {
+            # 'query': [{'role': 'user', 'content': query, 'loss': None}],
+            'query': query,
+            'positive_messages': [
+                 [{'role': 'user', 'content': passage_positive[i]}] for i in range(len(passage_positive))
+                ],
+            'negative_messages': [
+                 [{'role': 'user', 'content': all_neg[i]}] for i in range(len(all_neg))
+                ],
+            # 'label': 1.0
+        }
+        if len(row["negative_messages"]) == 0:
+            del row["negative_messages"]
+        return super().preprocess(row)
+
+
+register_dataset(
+    DatasetMeta(
+        dataset_path=os.path.dirname(__file__) + '/dataset_train.json',
+        dataset_name="my_local_dataset",
+        # subsets=[SubsetDataset('train', split=['train']), SubsetDataset('test', split=['test'])],
+        preprocess_func=CustomPreprocessor(),
+    ))
+
+if __name__ == '__main__':
+    # load_dataset returns train_dataset and val_dataset based on `split_dataset_ratio`
+    # Here, since we didn't pass `split_dataset_ratio` (defaults to 0), we take the first one (index 0)
+    dataset = load_dataset('my_local_dataset')[0]
+    test_dataset = load_dataset('swift/financial_classification:test')[0]
+    print(f'dataset[0]: {dataset[0]}')
+    print(f'test_dataset[0]: {test_dataset[0]}')

data/dataset/v11_dataset_hn/generated.py

@@ -0,0 +1,57 @@
+from swift.llm import ResponsePreprocessor, DatasetMeta, register_dataset, SubsetDataset, load_dataset
+from typing import Dict, Any
+import os
+
+
+class CustomPreprocessor(ResponsePreprocessor):
+    # def __init__(self, *, columns = None, **kwargs):
+    #     super().__init__(columns=columns, **kwargs)
+    #     self.num_all_negative = 0
+    def get_detailed_instruct(self, task_description: str, query: str) -> str:
+        return f'Instruct: {task_description}\nQuery:{query}'
+
+    def add_template(self, text):
+        task = 'Given a web search query, retrieve relevant passages that answer the query'
+        return self.get_detailed_instruct(task, text)
+
+    def preprocess(self, row: Dict[str, Any]) -> Dict[str, Any]:
+        query = self.add_template(row["query"])
+        passage_positive = row["passage_positive"]
+        passage_negative = row["passage_negative"]
+        passage_negative_random = row["passage_negative_random"]
+
+        all_neg = passage_negative + passage_negative_random
+        all_neg = list(set(all_neg))
+        # self.num_all_negative += len(all_neg)
+
+        row = {
+            # 'query': [{'role': 'user', 'content': query, 'loss': None}],
+            'query': query,
+            'positive_messages': [
+                 [{'role': 'user', 'content': passage_positive[i]}] for i in range(len(passage_positive))
+                ],
+            'negative_messages': [
+                 [{'role': 'user', 'content': all_neg[i]}] for i in range(len(all_neg))
+                ],
+            # 'label': 1.0
+        }
+        if len(row["negative_messages"]) == 0:
+            del row["negative_messages"]
+        return super().preprocess(row)
+
+
+register_dataset(
+    DatasetMeta(
+        dataset_path=os.path.dirname(__file__) + '/v11_dataset_hn.json',
+        dataset_name="v11_dataset_hn",
+        # subsets=[SubsetDataset('train', split=['train']), SubsetDataset('test', split=['test'])],
+        preprocess_func=CustomPreprocessor(),
+    ))
+
+if __name__ == '__main__':
+    # load_dataset returns train_dataset and val_dataset based on `split_dataset_ratio`
+    # Here, since we didn't pass `split_dataset_ratio` (defaults to 0), we take the first one (index 0)
+    dataset = load_dataset('v11_dataset_hn')[0]
+    test_dataset = load_dataset('swift/financial_classification:test')[0]
+    print(f'dataset[0]: {dataset[0]}')
+    print(f'test_dataset[0]: {test_dataset[0]}')

data/dataset/v11_generated/generated.py

@@ -0,0 +1,53 @@
+from swift.llm import ResponsePreprocessor, DatasetMeta, register_dataset, SubsetDataset, load_dataset
+from typing import Dict, Any
+import os
+
+
+class CustomPreprocessor(ResponsePreprocessor):
+    def get_detailed_instruct(self, task_description: str, query: str) -> str:
+        return f'Instruct: {task_description}\nQuery:{query}'
+
+    def add_template(self, text):
+        task = 'Given a web search query, retrieve relevant passages that answer the query'
+        return self.get_detailed_instruct(task, text)
+
+    def preprocess(self, row: Dict[str, Any]) -> Dict[str, Any]:
+        query = self.add_template(row["query"])
+        passage_positive = [row["document"]]
+        passage_negative = []
+        passage_negative_random = []
+
+        all_neg = passage_negative + passage_negative_random
+        all_neg = list(set(all_neg))
+
+        row = {
+            # 'query': [{'role': 'user', 'content': query, 'loss': None}],
+            'query': query,
+            'positive_messages': [
+                 [{'role': 'user', 'content': passage_positive[i]}] for i in range(len(passage_positive))
+                ],
+            'negative_messages': [
+                 [{'role': 'user', 'content': all_neg[i]}] for i in range(len(all_neg))
+                ],
+            # 'label': 1.0
+        }
+        if len(row["negative_messages"]) == 0:
+            del row["negative_messages"]
+        return super().preprocess(row)
+
+
+register_dataset(
+    DatasetMeta(
+        dataset_path=os.path.dirname(__file__) + '/v11_dataset.json',
+        dataset_name="v11_generated_dataset",
+        # subsets=[SubsetDataset('train', split=['train']), SubsetDataset('test', split=['test'])],
+        preprocess_func=CustomPreprocessor(),
+    ))
+
+if __name__ == '__main__':
+    # load_dataset returns train_dataset and val_dataset based on `split_dataset_ratio`
+    # Here, since we didn't pass `split_dataset_ratio` (defaults to 0), we take the first one (index 0)
+    dataset = load_dataset('v11_generated_dataset')[0]
+    test_dataset = load_dataset('swift/financial_classification:test')[0]
+    print(f'dataset[0]: {dataset[0]}')
+    print(f'test_dataset[0]: {test_dataset[0]}')

evaluation/evaluation.py

@@ -56,7 +56,7 @@ class CustomModel:
         **kwargs,
     ) -> np.ndarray:
 
-        embedding_url = "http://127.0.0.1:5000/embedding"
+        embedding_url = "http://127.0.0.1:5015/embedding"
         
         if prompt_type == None:
             template = "document"
@@ -89,6 +89,8 @@ def is_dataset_cached(dataset_name):
 
 def evaluate():
     model_name = "Qwen3-Embedding-0.6B"
+    # model_name = "KaLM-embedding-multilingual-mini-instruct-v2.5"
+    # model_name = "KaLM-Embedding-Gemma3-12B-2511"
     # model_name = "llama-embed-nemotron-8b"
     # model_name = "embeddinggemma-300m"
     model = CustomModel(model_name)

train/qwen/a.sh

@@ -3,6 +3,9 @@
 
 nproc_per_node=1
 
+# INFONCE_HARD_NEGATIVES=1 \
+# INFONCE_MASK_FAKE_NEGATIVE=True \
+
 MLFLOW_TRACKING_URI=http://0.0.0.0:5004 \
 INFONCE_USE_BATCH=True \
 CUDA_VISIBLE_DEVICES=0 \
@@ -16,8 +19,8 @@ swift sft \
     --lora_alpha 32 \
     --target_modules all-linear \
     --max_length 2048  \
-    --dataset v11_dataset_hn \
+    --dataset generated_250000_religous \
-    --custom_register_path $(pwd)/../../data/dataset/v11_dataset_hn/generated.py \
+    --custom_register_path $(pwd)/../../data/dataset/generated_250000_religous/generated.py \
     --split_dataset_ratio 0.005 \
     --eval_strategy steps \
     --output_dir output \

train/qwen/merge_model.py

@@ -33,8 +33,8 @@ 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/v17-20251202-223944/checkpoint-387"
+    peft_model_path = file_path + "/output/v23-20251214-111804/checkpoint-3632"
-    save_path = file_path + "/output/v17-20251202-223944/merged_checkpoint-387"
+    save_path = file_path + "/output/v23-20251214-111804/merged_checkpoint-3632"
     merge(base_model_path, peft_model_path, save_path)
 
     items = ["1_Pooling", "config_sentence_transformers.json", "merges.txt", "modules.json", "README.md", "tokenizer_config.json", "tokenizer.json",