热门标签
热门文章
- 1Java资料网盘分享_《java web从入门到精通》第一版光盘资源
- 2P1068 [NOIP2009 普及组] 分数线划定 Python
- 3pat甲级题库
- 4Kafka中那些巧妙的设计_kafka设计
- 5MapReduce场景应用_mapreduce的应用场景
- 6我们的系统检测到您的计算机网络中存在异常流量。此网页用于确认这些请求是由您而不是自动程序发出的。_我们的系统检测到您的计算机网络中存在异常流量。请稍后重新发送请求
- 7iOS-砸壳篇（两种砸壳方式）_ios砸壳工具
- 8Android打包系列 JAR反编译修改后重新打包_安卓 修改jar再打包
- 9学习神经网络大模型 电脑什么配置比较好？_跑大模型需要什么样电脑
- 10协议--QUIC理解_流量构成quic
当前位置: article > 正文
Qwen2（Qwen1.5阿里）微调实战qwen1.5-7b_qwen 本地微调如何配置验证集
作者：你好赵伟 | 2024-06-07 21:09:11
赞
踩
qwen 本地微调如何配置验证集
Qwen1.5-SFT
Qwen1.5-SFT(阿里, Ali), Qwen_Qwen1.5-2B-Chat/Qwen_Qwen1.5-7B-Chat微调(transformers)/LORA(peft)/推理
项目地址
- https://github.com/yongzhuo/qwen1.5-sft
- 默认数据类型为bfloat6
备注
1. 非常重要: weights要用bfloat16/fp32/tf32(第二版大模型基本共识), 不要用fp16, fp16会特别容易loss=NAN; 2. SFT最好还是像预训练那样, input/output都计算loss; 2. transformers需要4.37.0及以上; 3. qwen1.5, 0.5/1.8/4/7B模型的词典大小为151936(可以除以128), 14/72B模型的字典大小为152064(可以除以256); 多语言版本, 包含繁/简体; 4. qwen1.5网络架构同Llama; 5. prompt: 5.1 标准格式为: text_input + text_output + <|endoftext|> 5.2 prompt格式为: <|im_start|>system {text_prompt}<|im_end|> <|im_start|>user {text_input} <|im_end|> <|im_start|>assistant {text_output} <|im_end|> 6 微调输入输出: 输入："<|im_start|>system\n{text_prompt}<|im_end|>\n<|im_start|>user\n{text_input}\n<|im_end|>" 输出："<|im_start|>assistant\n{text_output}\n<|im_end|>" 7 推理输入输出(assistant\n放置位置不同): 输入："<|im_start|>system\n{text_prompt}<|im_end|>\n<|im_start|>user\n{text_input}\n<|im_end|>\n<|im_start|>assistant" 输出："{text_output}\n<|im_end|>" 8. 网络各层名称 ('base_model.model.model.embed_tokens.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.self_attn.q_proj.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.self_attn.q_proj.bias', torch.bfloat16, False) ('base_model.model.model.layers.31.self_attn.q_proj.lora_A.default.weight', torch.float32, True) ('base_model.model.model.layers.31.self_attn.q_proj.lora_B.default.weight', torch.float32, True) ('base_model.model.model.layers.31.self_attn.k_proj.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.self_attn.k_proj.bias', torch.bfloat16, False) ('base_model.model.model.layers.31.self_attn.k_proj.lora_A.default.weight', torch.float32, True) ('base_model.model.model.layers.31.self_attn.k_proj.lora_B.default.weight', torch.float32, True) ('base_model.model.model.layers.31.self_attn.v_proj.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.self_attn.v_proj.bias', torch.bfloat16, False) ('base_model.model.model.layers.31.self_attn.v_proj.lora_A.default.weight', torch.float32, True) ('base_model.model.model.layers.31.self_attn.v_proj.lora_B.default.weight', torch.float32, True) ('base_model.model.model.layers.31.self_attn.o_proj.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.mlp.gate_proj.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.mlp.up_proj.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.mlp.down_proj.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.input_layernorm.weight', torch.bfloat16, False) ('base_model.model.model.layers.31.post_attention_layernorm.weight', torch.bfloat16, False) ('base_model.model.model.norm.weight', torch.bfloat16, False) ('base_model.model.lm_head.weight', torch.bfloat16, False) 9. RuntimeError: unscale_() has already been called on this optimizer since the last update(). 微调语料太少导致的
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
环境配置
transformers>=4.37.0
torch>=1.11.0
safetensors>=0.4.1
accelerate==0.27.1
fsspec==2023.9.2
rouge==1.0.1
nltk==3.6.6
peft>=0.2.0
numpy
tqdm
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
微调
地址: qwen2_sft/ft_qwen2
配置: qwen2_sft/ft_qwen2/config.py
训练: python train.py
推理: python predict.py
验证: python evaluation.py
接口: python post_api.py
- 1
- 2
- 3
- 4
- 5
- 6
- 7
数据集-中文
- https://huggingface.co/datasets/JosephusCheung/GuanacoDataset
- https://huggingface.co/datasets/shareAI/shareGPT_cn
- https://huggingface.co/datasets/Mutonix/RefGPT-Fact
- https://huggingface.co/datasets/BAAI/COIG
- https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM
- https://github.com/carbonz0/alpaca-chinese-dataset
- https://github.com/LianjiaTech/BELLE
- https://github.com/PhoebusSi/Alpaca-CoT
- https://github.com/Hello-SimpleAI/chatgpt-comparison-detection
- https://github.com/yangjianxin1/Firefly
- https://github.com/XueFuzhao/InstructionWild
- https://github.com/OpenLMLab/MOSS
- https://github.com/thu-coai/Safety-Prompts
- https://github.com/LAION-AI/Open-Assistant
- https://github.com/TigerResearch/TigerBot
参考/感谢
- https://github.com/QwenLM/Qwen1.5
- https://github.com/google/gemma_pytorch
- https://huggingface.co/google/gemma-2b-it
- https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard
- https://github.com/THUDM/ChatGLM-6B
- https://github.com/THUDM/GLM
- https://github.com/tatsu-lab/stanford_alpaca
- https://github.com/LianjiaTech/BELLE
- https://github.com/huggingface/peft
- https://github.com/mymusise/ChatGLM-Tuning
- https://github.com/bojone/bert4keras
- trl
- math23k
微调日志-advgen
默认的bf16可微调, 最好也计算全部loss(同pt, input/output)
fp16损失函数就为nan, image_advgen_loss_nan_7b_3e4_constranr_bs4_fp16
fp16细调超参数偶尔训练成功但loss很高, image_advgen_dev_success_7b_2e5_cosine_bs2_fp16
推理日志-advgen
('base_model.model.base_model.model.model.layers.31.self_attn.k_proj.bias', torch.float16, False, tensor([ 0.1016, -0.0640, 0.1729, ..., -1.0312, -0.0028, 0.3477], device='cuda:0', dtype=torch.float16)) ('base_model.model.base_model.model.model.layers.31.self_attn.k_proj.lora_A.default.weight', torch.float16, False, tensor([[-0.0045, -0.0027, 0.0057, ..., 0.0131, 0.0074, -0.0049], [-0.0049, 0.0048, -0.0067, ..., 0.0118, 0.0080, -0.0105], [-0.0052, 0.0086, 0.0070, ..., -0.0102, -0.0005, 0.0166], ..., [-0.0097, -0.0158, -0.0116, ..., 0.0051, 0.0046, 0.0015], [-0.0148, 0.0001, -0.0152, ..., 0.0063, 0.0139, 0.0026], [-0.0092, 0.0036, 0.0003, ..., -0.0082, 0.0063, -0.0058]], device='cuda:0', dtype=torch.float16)) ('base_model.model.base_model.model.model.layers.31.self_attn.k_proj.lora_B.default.weight', torch.float16, False, tensor([[ 1.2102e-03, 1.9951e-03, -2.2488e-03, ..., 1.2598e-03, -1.4429e-03, -2.0466e-03], [-5.9128e-05, 4.4274e-04, 1.4553e-03, ..., -1.9350e-03, -8.8406e-04, 1.9855e-03], [ 1.5755e-03, 2.3174e-03, -2.6073e-03, ..., -4.1466e-03, -2.8076e-03, 6.1464e-04], ..., [-1.5554e-03, -3.0460e-03, 4.4136e-03, ..., 3.6411e-03, 6.6280e-04, 5.5790e-04], [-1.5869e-03, -1.6050e-03, -1.0443e-03, ..., 6.1393e-06, 1.7271e-03, -1.2022e-04], [-3.7169e-04, 6.4754e-04, 7.2241e-04, ..., 4.6730e-04, 7.0667e-04, -1.1092e-04]], device='cuda:0', dtype=torch.float16)) ...... ('base_model.model.base_model.model.lm_head.weight', torch.float16, False, tensor([[ 0.0095, -0.0017, -0.0134, ..., 0.0036, 0.0232, -0.0112], [ 0.0089, -0.0062, 0.0004, ..., -0.0079, 0.0175, -0.0098], [ 0.0055, 0.0182, 0.0088, ..., 0.0059, 0.0034, -0.0164], ..., [-0.0034, -0.0049, 0.0033, ..., -0.0022, -0.0066, 0.0031], [-0.0034, -0.0071, 0.0032, ..., -0.0022, -0.0066, 0.0013], [-0.0034, -0.0049, 0.0033, ..., -0.0016, -0.0066, 0.0020]], device='cuda:0', dtype=torch.float16)) trainable params: 0 || all params: 7733907456 || trainable%: 0.0 151645 151644 151645 1699 872 8948 {'instruction': '类型#裙*版型#宽松*版型#显瘦*颜色#黑色*图案#撞色*裙型#直筒裙*裙款式#拼接', 'input': '', 'output': ''} tensor([[151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 31905, 2, 102807, 9, 40301, 24300, 2, 109285, 9, 40301, 24300, 2, 99580, 102372, 9, 102284, 2, 104723, 9, 108108, 2, 101340, 38035, 9, 102807, 24300, 2, 73145, 106000, 102807, 9, 102807, 108756, 2, 100278, 29077, 151645, 198, 151644, 77091]], device='cuda:0') <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user 类型#裙*版型#宽松*版型#显瘦*颜色#黑色*图案#撞色*裙型#直筒裙*裙款式#拼接<|im_end|> <|im_start|>assistant宽松显瘦的黑色直筒裙是女性衣橱里必不可少的单品，这款裙子采用经典的黑色调，十分显瘦显腿长。裙摆处采用撞色的拼接设计，增添层次感的同时又不失时尚气息。裙型采用直筒裙的版型，更加修饰腿型，遮肉效果极佳。无论是职场还是休闲都能轻松驾驭，轻松展现女性的优雅气质。<|im_end|> 请输入: 类型#上衣*材质#棉*颜色#白色*风格#休闲*图案#刺绣*衣样式#衬衫*衣领型#尖领*衣袖长#七分袖 请稍等... ################################################################################################################################ {'instruction': '类型#上衣*材质#棉*颜色#白色*风格#休闲*图案#刺绣*衣样式#衬衫*衣领型#尖领*衣袖长#七分袖', 'input': '', 'output': ''} tensor([[151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 31905, 2, 17447, 99741, 9, 106149, 2, 102430, 9, 102284, 2, 102440, 9, 104040, 2, 104443, 9, 108108, 2, 100253, 103084, 9, 99741, 112453, 2, 113727, 9, 99741, 99213, 24300, 2, 100512, 99213, 9, 99741, 102885, 45861, 2, 99612, 17177, 102885, 151645, 198, 151644, 77091]], device='cuda:0') <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user 类型#上衣*材质#棉*颜色#白色*风格#休闲*图案#刺绣*衣样式#衬衫*衣领型#尖领*衣袖长#七分袖<|im_end|> <|im_start|>assistant这款白色刺绣休闲衬衫采用优质的纯棉材质，亲肤透气，穿着舒适亲肤。经典的尖领设计，修饰脸型，展现个性风格。七分袖的设计，既能凸显出修长的手型，又能带来一丝休闲的韵味。精致的刺绣图案，层次丰富，细节精致，凸显出独特的个性品味。这款衬衫无论是日常出街还是职场穿搭，都能展现出独特的休闲风格。<|im_end|> 27.06706738471985 请输入: 类型#裙*版型#显瘦*材质#牛仔布*风格#复古*风格#性感*图案#复古*裙型#牛仔裙*裙型#包臀裙*裙下摆#开叉*裙下摆#毛边*裙长#半身裙 请稍等... ################################################################################################################################ {'instruction': '类型#裙*版型#显瘦*材质#牛仔布*风格#复古*风格#性感*图案#复古*裙型#牛仔裙*裙型#包臀裙*裙下摆#开叉*裙下摆#毛边*裙长#半身裙', 'input': '', 'output': ''} tensor([[151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 31905, 2, 102807, 9, 40301, 24300, 2, 99580, 102372, 9, 106149, 2, 100664, 102437, 51827, 9, 104040, 2, 111455, 9, 104040, 2, 110803, 9, 108108, 2, 111455, 9, 102807, 24300, 2, 100664, 102437, 102807, 9, 102807, 24300, 2, 67279, 117318, 102807, 9, 102807, 16872, 100805, 2, 29767, 102940, 9, 102807, 16872, 100805, 2, 99676, 99319, 9, 102807, 45861, 2, 99369, 95256, 102807, 151645, 198, 151644, 77091]], device='cuda:0') <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user 类型#裙*版型#显瘦*材质#牛仔布*风格#复古*风格#性感*图案#复古*裙型#牛仔裙*裙型#包臀裙*裙下摆#开叉*裙下摆#毛边*裙长#半身裙<|im_end|> <|im_start|>assistant这款牛仔裙采用了复古风格的设计，十分的复古又有性感的感觉。牛仔布的材质十分的柔软亲肤，而且牛仔的风格十分的显瘦，很好的修饰身材线条，凸显出女性的优雅气质。这款半身裙的设计十分的别致，裙摆处采用开叉的设计，十分的个性时尚，凸显出女性的个性魅力。裙摆处采用毛边的设计，十分的个性时尚，而且层次感十足，凸显出女性的优雅气质。而且这款半裙还是包臀裙的设计，很好的修饰出女性的身材线条，凸显出女性的优雅气质 14.573514223098755 请输入: 类型#裤*材质#牛仔布*颜色#白色*风格#简约*图案#线条*裤长#短裤*裤型#阔腿裤*裤腰型#高腰*裤口#毛边 请稍等... ################################################################################################################################ {'instruction': '类型#裤*材质#牛仔布*颜色#白色*风格#简约*图案#线条*裤长#短裤*裤型#阔腿裤*裤腰型#高腰*裤口#毛边', 'input': '', 'output': ''} tensor([[151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 31905, 2, 102693, 9, 106149, 2, 100664, 102437, 51827, 9, 102284, 2, 102440, 9, 104040, 2, 109962, 9, 108108, 2, 108236, 9, 102693, 45861, 2, 99534, 102693, 9, 102693, 24300, 2, 100607, 100447, 102693, 9, 102693, 102113, 24300, 2, 44636, 102113, 9, 102693, 39426, 2, 99676, 99319, 151645, 198, 151644, 77091]], device='cuda:0') <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user 类型#裤*材质#牛仔布*颜色#白色*风格#简约*图案#线条*裤长#短裤*裤型#阔腿裤*裤腰型#高腰*裤口#毛边<|im_end|> <|im_start|>assistant这款牛仔裤的设计风格十分简约，线条流畅，十分显瘦，而且高腰的设计十分拉长腿型，显高显瘦。裤子采用了白色的色系，十分清新甜美，而且十分百搭。阔腿裤的版型十分修饰腿型，十分显腿长，而且十分遮肉显瘦。裤脚采用了毛边的设计，十分复古有个性，而且十分甜美可爱。这款牛仔短裤十分百搭，而且十分显腿长，而且十分显瘦，而且十分甜美甜美，十分凸显青春活力的气息。<|im_end|> 22.39077615737915 请输入: 类型#裤*版型#显瘦*材质#牛仔布*颜色#深蓝色*裤腰型#高腰 请稍等... ################################################################################################################################ {'instruction': '类型#裤*版型#显瘦*材质#牛仔布*颜色#深蓝色*裤腰型#高腰', 'input': '', 'output': ''} tensor([[151644, 8948, 198, 2610, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 31905, 2, 102693, 9, 40301, 24300, 2, 99580, 102372, 9, 106149, 2, 100664, 102437, 51827, 9, 102284, 2, 99194, 105681, 9, 102693, 102113, 24300, 2, 44636, 102113, 151645, 198, 151644, 77091]], device='cuda:0') <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user 类型#裤*版型#显瘦*材质#牛仔布*颜色#深蓝色*裤腰型#高腰<|im_end|> <|im_start|>assistant这款牛仔裤采用深蓝色的色调，显得沉稳大方，同时也十分的显瘦。高腰的版型更是修饰了腿型，显高显腿长。牛仔布的材质十分的舒适亲肤，而且牛仔的质地也能凸显出个性的感觉，十分的青春活力。无论是日常的休闲搭配，还是职场的职场风格，都能很好的展现出女性的优雅知性。<|im_end|> 8.462041139602661
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
- 58
- 59
- 60
- 61
- 62
- 63
- 64
- 65
- 66
- 67
- 68
- 69
- 70
- 71
- 72
- 73
- 74
- 75
- 76
- 77
- 78
- 79
- 80
- 81
- 82
- 83
- 84
- 85
- 86
- 87
- 88
- 89
- 90
- 91
- 92
- 93
- 94
- 95
- 96
- 97
- 98
- 99
- 100
- 101
- 102
- 103
- 104
- 105
- 106
- 107
- 108
- 109
- 110
- 111
- 112
- 113
- 114
- 115
- 116
- 117
- 118
- 119
- 120
- 121
- 122
- 123
- 124
- 125
- 126
- 127
.<|im_end|>
<|im_start|>user
类型#裤版型#显瘦材质#牛仔布颜色#深蓝色裤腰型#高腰<|im_end|>
<|im_start|>assistant这款牛仔裤采用深蓝色的色调，显得沉稳大方，同时也十分的显瘦。高腰的版型更是修饰了腿型，显高显腿长。牛仔布的材质十分的舒适亲肤，而且牛仔的质地也能凸显出个性的感觉，十分的青春活力。无论是日常的休闲搭配，还是职场的职场风格，都能很好的展现出女性的优雅知性。<|im_end|>
8.462041139602661
希望对你有所帮助!
- 1
- 2
- 3
声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/你好赵伟/article/detail/687030
推荐阅读
相关标签
