haotian/unsloth
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

微调unsloth系列模型

Browse Source
This commit is contained in:
haotian 2025-07-15 11:49:46 +08:00
parent 1f0a420d7d
commit d14fe58e58
3 changed files with 10 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
003加载自己的数据集微调.py
View File

@ -23,18 +23,23 @@ dataset = dataset.map(to_chatml)
# print("\n", dataset[0]) # print("\n", dataset[0])
# 加载预训练模型 # 加载预训练模型
model, tokenizer = FastLanguageModel.from_pretrained( model, tokenizer = FastLanguageModel.from_pretrained(
# model_name = "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", # model_name = "unsloth/Qwen3-1.7B-unsloth-bnb-4bit",
# model_name = "unsloth/Qwen3-8B-unsloth-bnb-4bit", model_name = "unsloth/Qwen3-8B-unsloth-bnb-4bit",
model_name = "deepseek-ai/DeepSeek-V2-Lite", # model_name = "deepseek-ai/DeepSeek-V2-Lite",
# model_name = "unsloth/Qwen3-4B-unsloth-bnb-4bit", # model_name = "unsloth/Qwen3-4B-unsloth-bnb-4bit",
trust_remote_code=True, # ✅ 允許遠端自訂程式碼, deepseek 系列模型使用 trust_remote_code=True, # ✅ 允許遠端自訂程式碼, deepseek 系列模型使用
max_seq_length = 2048, # Context length - can be longer, but uses more memory max_seq_length = 2048, # Context length - can be longer, but uses more memory
# max_seq_length = 512, # Context length - can be longer, but uses more memory
load_in_4bit = True, # 4bit uses much less memory , 启用QLoRA load_in_4bit = True, # 4bit uses much less memory , 启用QLoRA
load_in_8bit = False, # A bit more accurate, uses 2x memory load_in_8bit = False, # A bit more accurate, uses 2x memory
full_finetuning = False, # We have full finetuning now! full_finetuning = False, # We have full finetuning now!
# token = "hf_...", # use one if using gated models # token = "hf_...", # use one if using gated models
) )
model = FastLanguageModel.get_peft_model( model = FastLanguageModel.get_peft_model(
@ -100,7 +105,7 @@ _ = model.generate(
# model.cpu() # model.cpu()
model.save_pretrained_gguf( model.save_pretrained_gguf(
"DeepSeek-V2-Lite", "Qwen3-8B",
tokenizer, tokenizer,
# quantization_method="q4_k_m", # 或 "q8_0" # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m", # quantization_method="q4_k_m", # 或 "q8_0" # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m",
maximum_memory_usage=0.7 # 限制使用 GPU 显存为总容量的 50% maximum_memory_usage=0.7 # 限制使用 GPU 显存为总容量的 50%

2
unsloth_compiled_cache/UnslothAlignPropTrainer.py
View File

@ -130,7 +130,7 @@ class UnslothAlignPropConfig(AlignPropConfig):
) )
def __init__( def __init__(
self, self,
exp_name = '003加载自己的数据集微调', exp_name = '003微调deepseek',
run_name = '', run_name = '',
seed = 3407, seed = 3407,
log_with = None, log_with = None,

2
unsloth_compiled_cache/UnslothDDPOTrainer.py
View File

@ -146,7 +146,7 @@ class UnslothDDPOConfig(DDPOConfig):
) )
def __init__( def __init__(
self, self,
exp_name = '003加载自己的数据集微调', exp_name = '003微调deepseek',
run_name = '', run_name = '',
seed = 3407, seed = 3407,
log_with = None, log_with = None,