feat/add-support-qwen3.5

demo.py

@@ -5,7 +5,7 @@
 from peft import LoraConfig
 from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
 from trl import SFTTrainer, SFTConfig
-
+from datasets import Dataset
 from dataset import SFTDataCollator, SFTDataset
 from utils.constants import model2template
 
@@ -53,7 +53,6 @@ def train_lora(
         optim="paged_adamw_8bit",
         remove_unused_columns=False,
         num_train_epochs=training_args.num_train_epochs,
-        max_seq_length=context_length,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_id,
@@ -74,10 +73,12 @@ def train_lora(
         template=model2template[model_id],
     )
 
+    dataset_dict = [dataset[i] for i in range(len(dataset))]
+    hf_dataset = Dataset.from_list(dataset_dict)
     # Define trainer
     trainer = SFTTrainer(
         model=model,
-        train_dataset=dataset,
+        train_dataset=hf_dataset,
         args=training_args,
         peft_config=lora_config,
         data_collator=SFTDataCollator(tokenizer, max_seq_length=context_length),
@@ -108,7 +109,7 @@ def train_lora(
     )
 
     # Set model ID and context length
-    model_id = "Qwen/Qwen1.5-0.5B"
+    model_id = "Qwen/Qwen3.5-2B"
     context_length = 2048
 
     # Start LoRA fine-tuning

requirements.txt

@@ -1,7 +1,8 @@
+pyyaml
 torch>=1.13.1
-transformers>=4.43.0,<=4.45.0
-peft>=0.10.0,<=0.13.2
-loguru
-trl>=0.9.3,<=0.9.6
+transformers==5.3.0
+peft==0.18.1
+loguru>=0.6.0
+huggingface-hub==1.5.0
+trl>=0.20.0,<=0.29.1
 bitsandbytes
-pyyaml

utils/constants.py

@@ -9,19 +9,37 @@
 }
 
 model2template = {
-    "Qwen/Qwen2.5-0.5B-Instruct": qwen_template,
-    "Qwen/Qwen2.5-1.5B-Instruct": qwen_template,
-    "Qwen/Qwen2.5-7B-Instruct": qwen_template,
+    "Qwen/Qwen3.5-0.8B": qwen_template,
+    "Qwen/Qwen3.5-0.8B-Base": qwen_template,
+    "Qwen/Qwen3.5-2B": qwen_template,
+    "Qwen/Qwen3.5-2B-Base": qwen_template,
+    "Qwen/Qwen3.5-4B": qwen_template,
+    "Qwen/Qwen3.5-4B-Base": qwen_template,
+    "Qwen/Qwen3.5-9B": qwen_template,
+    "Qwen/Qwen3.5-9B-Base": qwen_template,
+    "Qwen/Qwen3.5-27B": qwen_template,
 }
 
 model2size = {
-    "Qwen/Qwen2.5-0.5B-Instruct": 494_000_000,
-    "Qwen/Qwen2.5-1.5B-Instruct": 1_540_000_000,
-    "Qwen/Qwen2.5-7B-Instruct": 7_620_000_000,
+    "Qwen/Qwen3.5-0.8B": 853_000_000,
+    "Qwen/Qwen3.5-0.8B-Base": 853_000_000,
+    "Qwen/Qwen3.5-2B": 2_213_000_000,
+    "Qwen/Qwen3.5-2B-Base": 2_213_000_000,
+    "Qwen/Qwen3.5-4B": 4_539_000_000,
+    "Qwen/Qwen3.5-4B-Base": 4_539_000_000,
+    "Qwen/Qwen3.5-9B": 8_392_000_000,
+    "Qwen/Qwen3.5-9B-Base": 8_392_000_000,
+    "Qwen/Qwen3.5-27B": 26_085_000_000,
 }
 
 model2base_model = {
-    "Qwen/Qwen2.5-0.5B-Instruct": "qwen1.5",
-    "Qwen/Qwen2.5-1.5B-Instruct": "qwen1.5",
-    "Qwen/Qwen2.5-7B-Instruct": "qwen1.5",
+    "Qwen/Qwen3.5-0.8B": "qwen3.5",
+    "Qwen/Qwen3.5-0.8B-Base": "qwen3.5",
+    "Qwen/Qwen3.5-2B": "qwen3.5",
+    "Qwen/Qwen3.5-2B-Base": "qwen3.5",
+    "Qwen/Qwen3.5-4B": "qwen3.5",
+    "Qwen/Qwen3.5-4B-Base": "qwen3.5",
+    "Qwen/Qwen3.5-9B": "qwen3.5",
+    "Qwen/Qwen3.5-9B-Base": "qwen3.5",
+    "Qwen/Qwen3.5-27B": "qwen3.5",
 }