LLaMA3（Meta）微调SFT实战Meta-Llama-3-8B-Instruct

2024-04-30 14:04
文章标签 meta 实战 微调 llama 8b instruct sft llama3

本文主要是介绍LLaMA3（Meta）微调SFT实战Meta-Llama-3-8B-Instruct，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

LlaMA3-SFT

LlaMA3-SFT, Meta-Llama-3-8B/Meta-Llama-3-8B-Instruct微调(transformers)/LORA(peft)/推理

项目地址

  • https://github.com/yongzhuo/LLaMA3-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4.40.0及以上;
3. llama3, 模型的词典大小为128256, 使用tiktoken;
4. llama3网络架构同Llama2, 使用GQA/MQA等加速; 
5. prompt:5.1 标准格式为: 
text_input + text_output5.2 prompt格式为: 
<|begin_of_text|><|start_header_id|>system<|end_header_id|>{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>{text_input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>{text_output}<|eot_id|>
6 微调输入输出:输入："<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{text_input}<|eot_id|>"输出："<|start_header_id|>assistant<|end_header_id|>\n\n{text_output}<|eot_id|>"
7 推理输入输出(assistant\n放置位置不同):输入："<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{text_input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"输出："{text_output}<|eot_id|>"
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.k_proj.weight', torch.bfloat16, False)
('base_model.model.model.layers.31.self_attn.v_proj.weight', torch.bfloat16, False)
('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

环境配置

transformers>=4.44.0
torch>=1.13.0
safetensors>=0.4.1
accelerate==0.27.1
fsspec==2023.9.2
rouge==1.0.1
nltk==3.6.6
peft>=0.2.0
tiktoken
numpy
tqdm

微调

地址: llama3_sft/ft_llama3配置: llama3_sft/ft_llama3/config.py
训练: python train.py
推理: python predict.py
验证: python evaluation.py
接口: python post_api.py

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/meta-llama/llama3
  • 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微调, 计算全部loss(同pt, input/output), loss太低

在这里插入图片描述在这里插入图片描述

推理日志-advgen(SFT), 感觉学到了一些东西, 又没有学会

('base_model.model.base_model.model.model.layers.31.self_attn.q_proj.weight', torch.bfloat16, False, tensor([[ 0.0078, -0.0035,  0.0075,  ..., -0.0007, -0.0016, -0.0020],[ 0.0036, -0.0019,  0.0140,  ..., -0.0161, -0.0037,  0.0012],[ 0.0080,  0.0125, -0.0103,  ..., -0.0005, -0.0267,  0.0027],...,[-0.0061,  0.0166,  0.0148,  ..., -0.0105, -0.0165,  0.0242],[-0.0181, -0.0244,  0.0100,  ..., -0.0334,  0.0020, -0.0227],[-0.0042, -0.0181,  0.0114,  ...,  0.0027, -0.0004, -0.0002]],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.self_attn.q_proj.lora_A.default.weight', torch.float32, False, tensor([[-0.0188,  0.0052, -0.0048,  ..., -0.0073,  0.0091,  0.0070],[-0.0117, -0.0048, -0.0020,  ...,  0.0009,  0.0169,  0.0139],[ 0.0011, -0.0023, -0.0166,  ..., -0.0071, -0.0082,  0.0082],...,[ 0.0127, -0.0045, -0.0137,  ..., -0.0146,  0.0114, -0.0145],[ 0.0038, -0.0148,  0.0001,  ..., -0.0172, -0.0139,  0.0062],[-0.0056,  0.0132,  0.0115,  ...,  0.0100,  0.0127,  0.0004]]))
('base_model.model.base_model.model.model.layers.31.self_attn.q_proj.lora_B.default.weight', torch.float32, False, tensor([[ 6.0165e-04, -1.8864e-04,  5.4384e-04,  ...,  3.9676e-04,-1.9826e-04, -7.4172e-05],[ 2.4563e-03, -2.3599e-03,  2.0984e-03,  ...,  1.9662e-03,1.5514e-03, -1.8005e-03],[-4.9125e-03,  4.7831e-03, -4.6263e-03,  ..., -4.3837e-03,-3.4190e-03,  4.2660e-03],...,[ 1.2281e-05,  5.1583e-04, -1.4735e-04,  ...,  1.3617e-04,-9.5673e-04,  8.3033e-04],[-1.5139e-03,  3.0022e-03, -2.4098e-03,  ..., -1.9860e-03,-2.3010e-03,  2.5711e-03],[ 1.6433e-03, -2.0675e-03,  2.2618e-03,  ...,  1.4840e-03,1.1958e-03, -1.3300e-03]]))
('base_model.model.base_model.model.model.layers.31.self_attn.k_proj.weight', torch.bfloat16, False, tensor([[ 0.0126, -0.0269,  0.0297,  ...,  0.0481, -0.0057,  0.0061],[-0.0262, -0.0154, -0.0427,  ..., -0.0217, -0.0513,  0.0092],[ 0.0098,  0.0058,  0.0157,  ..., -0.0054, -0.0025, -0.0024],...,[-0.0198,  0.0408, -0.0136,  ..., -0.0173, -0.0432,  0.0082],[ 0.0312,  0.0239,  0.0247,  ..., -0.0171, -0.0284, -0.0432],[-0.0096,  0.0043,  0.0142,  ..., -0.0164, -0.0386,  0.0206]],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.self_attn.k_proj.lora_A.default.weight', torch.float32, False, tensor([[ 0.0137, -0.0115,  0.0137,  ..., -0.0069, -0.0056, -0.0032],[-0.0117, -0.0096, -0.0165,  ..., -0.0014,  0.0045, -0.0028],[ 0.0072, -0.0115,  0.0109,  ..., -0.0169, -0.0165,  0.0081],...,[-0.0194,  0.0073, -0.0093,  ...,  0.0050, -0.0120,  0.0028],[-0.0064, -0.0142, -0.0112,  ..., -0.0126,  0.0087,  0.0097],[-0.0108,  0.0042, -0.0046,  ...,  0.0105, -0.0111, -0.0107]]))
('base_model.model.base_model.model.model.layers.31.self_attn.k_proj.lora_B.default.weight', torch.float32, False, tensor([[ 0.0024, -0.0033, -0.0026,  ..., -0.0024, -0.0035, -0.0017],[ 0.0016, -0.0022, -0.0032,  ..., -0.0017, -0.0023,  0.0015],[-0.0028,  0.0032,  0.0029,  ...,  0.0027,  0.0034,  0.0017],...,[-0.0015, -0.0001,  0.0005,  ...,  0.0001,  0.0006, -0.0047],[ 0.0022,  0.0023, -0.0005,  ..., -0.0007, -0.0013,  0.0028],[-0.0003, -0.0020, -0.0002,  ..., -0.0003, -0.0003, -0.0021]]))
('base_model.model.base_model.model.model.layers.31.self_attn.v_proj.weight', torch.bfloat16, False, tensor([[ 0.0172,  0.0181, -0.0032,  ..., -0.0250, -0.0186,  0.0119],[-0.0042,  0.0047, -0.0134,  ..., -0.0077,  0.0078, -0.0031],[-0.0422, -0.0249, -0.0136,  ..., -0.0295,  0.0302, -0.0123],...,[ 0.0076, -0.0002, -0.0001,  ...,  0.0093, -0.0047, -0.0044],[-0.0014, -0.0033, -0.0076,  ..., -0.0109, -0.0012, -0.0215],[-0.0232,  0.0025,  0.0060,  ...,  0.0068, -0.0093, -0.0073]],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.self_attn.v_proj.lora_A.default.weight', torch.float32, False, tensor([[ 0.0090,  0.0004, -0.0076,  ...,  0.0153,  0.0041, -0.0103],[ 0.0138,  0.0015, -0.0053,  ..., -0.0128,  0.0011, -0.0022],[-0.0111, -0.0032,  0.0006,  ..., -0.0077,  0.0128,  0.0151],...,[-0.0029,  0.0131,  0.0163,  ..., -0.0044, -0.0083, -0.0115],[ 0.0041,  0.0133,  0.0181,  ...,  0.0027,  0.0083,  0.0050],[ 0.0024, -0.0063,  0.0042,  ..., -0.0021, -0.0047,  0.0007]]))
('base_model.model.base_model.model.model.layers.31.self_attn.v_proj.lora_B.default.weight', torch.float32, False, tensor([[-3.0396e-04, -1.8482e-04, -4.5843e-04,  ..., -7.0217e-05,-7.8356e-04, -3.6097e-04],[-6.1859e-04, -1.8265e-04,  3.9717e-04,  ..., -1.0677e-03,-1.9114e-03, -1.1090e-03],[-3.2717e-04,  5.8097e-04,  2.8924e-04,  ...,  2.1774e-04,2.1301e-04, -4.1886e-04],...,[-1.9359e-03, -1.4303e-03, -1.5492e-03,  ..., -1.3788e-03,-1.8295e-03, -2.1401e-03],[-6.3756e-04, -5.9278e-04, -4.7388e-04,  ..., -4.0911e-04,-2.1048e-04, -5.5398e-04],[-1.9948e-03, -2.3558e-03, -2.5491e-03,  ..., -1.5711e-03,-1.9010e-03, -2.4418e-03]]))
('base_model.model.base_model.model.model.layers.31.self_attn.o_proj.weight', torch.bfloat16, False, tensor([[ 0.0198, -0.0121, -0.0223,  ...,  0.0117, -0.0031,  0.0131],[ 0.0003, -0.0032, -0.0046,  ...,  0.0013, -0.0013,  0.0123],[-0.0073,  0.0129, -0.0085,  ..., -0.0016,  0.0074, -0.0052],...,[ 0.0015,  0.0116, -0.0047,  ..., -0.0004, -0.0016,  0.0125],[-0.0167,  0.0030,  0.0166,  ..., -0.0014, -0.0126,  0.0087],[ 0.0172, -0.0017,  0.0156,  ..., -0.0197,  0.0104, -0.0012]],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.mlp.gate_proj.weight', torch.bfloat16, False, tensor([[-0.0142,  0.0153, -0.0243,  ...,  0.0134,  0.0041,  0.0069],[-0.0006,  0.0083,  0.0076,  ...,  0.0160,  0.0080,  0.0121],[-0.0128,  0.0200, -0.0142,  ...,  0.0060, -0.0074, -0.0006],...,[ 0.0054, -0.0041,  0.0251,  ...,  0.0022,  0.0177,  0.0177],[-0.0178, -0.0178,  0.0045,  ...,  0.0131, -0.0084, -0.0108],[-0.0269, -0.0266, -0.0234,  ...,  0.0366,  0.0393,  0.0086]],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.mlp.up_proj.weight', torch.bfloat16, False, tensor([[-1.6357e-02, -2.5513e-02, -6.4087e-03,  ...,  1.8555e-02,-7.5684e-03,  9.8877e-03],[ 6.4697e-03,  1.2436e-03,  1.2390e-02,  ...,  7.2937e-03,3.4668e-02, -3.9673e-03],[ 4.4556e-03,  2.2583e-03,  5.8594e-03,  ..., -8.6060e-03,1.4465e-02, -2.1484e-02],...,[ 2.5757e-02,  1.9836e-03,  1.8677e-02,  ...,  5.1880e-03,1.0864e-02,  9.9659e-05],[-3.5400e-03,  1.4221e-02, -8.6060e-03,  ...,  6.6833e-03,1.0071e-03,  1.2512e-02],[ 9.5825e-03, -1.3489e-02, -1.3885e-03,  ..., -4.4861e-03,-3.0823e-03, -9.3384e-03]], dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.mlp.down_proj.weight', torch.bfloat16, False, tensor([[-0.0334,  0.0152,  0.0194,  ...,  0.0025,  0.0114,  0.0013],[-0.0050, -0.0064,  0.0026,  ..., -0.0042,  0.0032, -0.0086],[-0.0094, -0.0220, -0.0214,  ...,  0.0029, -0.0057, -0.0084],...,[ 0.0125,  0.0017, -0.0042,  ..., -0.0166, -0.0077,  0.0110],[-0.0128,  0.0073,  0.0140,  ..., -0.0193,  0.0125,  0.0109],[-0.0074, -0.0139,  0.0104,  ...,  0.0105,  0.0035,  0.0027]],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.input_layernorm.weight', torch.bfloat16, False, tensor([0.4629, 0.3984, 0.4434,  ..., 0.4277, 0.3887, 0.2988],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.layers.31.post_attention_layernorm.weight', torch.bfloat16, False, tensor([0.5273, 0.4883, 0.5078,  ..., 0.5273, 0.4863, 0.4238],dtype=torch.bfloat16))
('base_model.model.base_model.model.model.norm.weight', torch.bfloat16, False, tensor([2.6562, 2.5781, 2.6094,  ..., 2.5938, 2.2656, 2.5156],dtype=torch.bfloat16))
('base_model.model.base_model.model.lm_head.weight', torch.bfloat16, False, tensor([[ 9.8267e-03,  1.7456e-02,  3.6926e-03,  ...,  5.7983e-04,-1.6968e-02, -1.0986e-02],[-6.5613e-03,  1.1719e-02,  1.1536e-02,  ..., -1.0193e-02,9.3994e-03, -1.4496e-03],[ 1.4160e-02,  9.4604e-03,  7.4463e-03,  ..., -1.9379e-03,-5.7678e-03, -1.4771e-02],...,[-3.7231e-03, -1.3065e-04,  7.5684e-03,  ...,  9.6798e-05,7.5989e-03,  6.0425e-03],[-3.7231e-03, -1.3065e-04,  7.5684e-03,  ...,  9.6798e-05,7.5989e-03,  6.0425e-03],[-3.7231e-03, -1.3065e-04,  7.5684e-03,  ...,  9.6798e-05,7.5989e-03,  6.0425e-03]], dtype=torch.bfloat16))
trainable params: 0 || all params: 8039698432 || trainable%: 0.0
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
128009
128000
128009
198
882
78191
{'content': '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'}
tensor([[128000, 128006,   9125, 128007,    271,   2675,    527,    264,  11190,18328,     11,    220,  15225,  11883,  99337,  33014, 108891, 113925,13, 128009, 128006,    882, 128007,    271,  33005,      2,  70892,247,      9,  41401,  25287,      2, 118188, 104500,      9,  41401,25287,      2, 105593, 114431,     99,      9, 124510,  39135,      2,57752,  39135,      9,  29129,  81742,      2,  58843,    252,  39135,9,  70892,    247,  25287,      2,  74245, 127946,  70892,    247,9,  70892,    247,  69253,  29430,      2, 125973,  30177, 128009,128006,  78191, 128007,    271]])
<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>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<|eot_id|><|start_header_id|>assistant<|end_header_id|>根据你的描述，这是一件黑色撞色拼接直筒裙，裙型宽松，版型显瘦，裙款式拼接。<|eot_id|>617.0838704109192
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...
################################################################################################################################
{'content': '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'}
tensor([[128000, 128006,   9125, 128007,    271,   2675,    527,    264,  11190,18328,     11,    220,  15225,  11883,  99337,  33014, 108891, 113925,13, 128009, 128006,    882, 128007,    271,  33005,      2,  70892,97,      9, 103234, 103706,      2, 109022, 120248,  52927,      9,124510,  39135,      2, 101828,  39135,      9, 103125,  35083,      2,99337,  95337,      9,  29129,  81742,      2,  44368,  40089,      9,70892,     97,  46961,      2, 106649,  70892,     97,      9,  70892,97,  25287,      2,  33443,    242, 122072,  70892,     97,      9,70892,     97, 115519,  25287,      2,  45736, 115519,      9,  70892,97,  40526,      2, 105811, 103377, 128009, 128006,  78191, 128007,271]])
<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>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<|eot_id|><|start_header_id|>assistant<|end_header_id|>根据您的描述，这是一双简约的短裤，使用牛仔布，白色为主色，线条图案，阔腿裤型，高腰，毛边裤口。<|eot_id|>

推理日志-advgen(SFT), 修改prompt微调, 指令遵循得不是太好, 感觉还是中文预训练语料太少了吧

<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请根据输入生成创意广告, 用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>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<|eot_id|><|start_header_id|>assistant<|end_header_id|>Here's a creative advertisement based on your input:**Headline:** 撞色撞焦点！黑色撞色拼接直筒裙**Body:**你准备好撞焦点了吗？我们的黑色撞色拼接直筒裙将让你在人群中脱颖而出！宽松版型，显瘦效果，让你显得更加苗条迷人。撞色图案，黑色拼接，更加增加裙子的时尚感和吸引力。这不是一件裙子，而是一件生活方式。穿上我们的黑色撞色拼接直筒裙，你将成为焦点，吸引所有人的目光。**Call to Action:** 快来试穿我们的黑色撞色拼接直筒裙，体验一下撞焦点的感觉！**Visuals:** A stunning image of a model wearing the black and color-blocked straight-cut skirt, posing confidently with a bold expression. The skirt is designed to accentuate her curves, and the color-blocked pattern adds a touch of edginess to the overall look. The background is a dark, sleek cityscape, emphasizing the
###################################################################
<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请根据输入生成创意广告, 用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>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<|eot_id|><|start_header_id|>assistant<|end_header_id|>Here's a creative ad for the specified type of pants:**Title:** 时尚简约，阔腿裤，白色牛仔布短裤**Image:** A stylish photo of a person wearing the white cowboy-style short pants with a simple and minimalist design. The pants have a wide leg and a high waist, with a subtle stitching pattern along the seams.**Headline:** 简约 yet 时尚，阔腿裤白色牛仔布短裤，让你展现自己的个性！**Body:** 不需要很多花哨的设计，阔腿裤白色牛仔布短裤就足以让你展现自己的个性！这款简约的短裤，以牛仔布为材料，白色为主色调，线条简洁，给人一种简洁 yet 时尚的感觉。无论是街头、工作或休闲，这款短裤都能满足你的需求，让你展现自己的个性！**Call-to-Action:** 快来抢购这款简约 yet 时尚的短裤，展现自己的个性！**Target Audience:** Young adults who value simplicity and individuality
################################################################################
<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请根据输入生成创意广告, 用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>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<|eot_id|><|start_header_id|>assistant<|end_header_id|>Here's a creative advertisement for the product:**Title:** 升级你的夏日风采！【简约牛仔短裤】**Image:** A stylish model wearing the white cowhide short pants with a high waist and wide legs, posing in a casual yet chic way. The pants have a subtle texture and a raw edge hem, adding a touch of edginess to the overall look.**Headline:** 升级你的夏日风采！简单 yet chic，简约牛仔短裤让你夏日更加精彩！**Body Text:** 夏日到了，时尚也到了！简约牛仔短裤是夏日的必备！高腰设计，阔腿裤造型，毛边裤口，简约牛仔短裤让你夏日更加精彩！不仅舒适还非常时尚，简约牛仔短裤是你的夏日必备！**Call-to-Action:** 快来获取你的简约牛仔短裤，升级你的夏日风采！This advertisement highlights the key features of the product, such as the high waist, wide legs, and raw edge hem, while also emphasizing its comfort and stylishness

推理日志-advgen(原始Instruct)

<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>类型#裙*材质#天丝*裙长#连衣裙*裙袖长#长袖<|eot_id|><|start_header_id|>assistant<|end_header_id|>Based on your input, I understand that you are looking for information about a dress with the following specifications:* Type: 连衣裙 (a-line dress)
* Material: 天丝 (silk)
* Length: 不指定 (no specific length mentioned, but since it's a 连衣裙, it's likely to be knee-length or longer)
* Sleeve length: 长袖 (long sleeves)Here's a summary of your requirements:* A-line dress made of silk with long sleeves and a knee-length or longer hem.If you have any further questions or would like to know more about this type of dress, feel free to ask! 😊<|eot_id|>
#############################################################
<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>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<|eot_id|><|start_header_id|>assistant<|end_header_id|>Based on your input, I can help you with the following information:* 类型：裙 (type: skirt)
* 版型：宽松 (style: loose-fitting)
* 版型：显瘦 (style: slimming)
* 颜色：黑色 (color: black)
* 图案：撞色 (pattern: contrast color)
* 裙型：直筒裙 (skirt type: straight skirt)
* 裙款式：拼接 (skirt style: patchwork)So, the overall description is: a black straight skirt with a loose-fitting and slimming design, featuring a contrast color pattern through patchwork.<|eot_id|>
#############################################################
<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a helpful assistant, 请用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>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<|eot_id|><|start_header_id|>assistant<|end_header_id|>您所描述的裤子是：* 材质：牛仔布（denim）
* 颜色：白色（white）
* 风格：简约（simple）
* 图案：线条（striped）
* 裤长：短裤（short）
* 裤型：阔腿裤（wide-leg）
* 裤腰型：高腰（high-waisted）
* 裤口：毛边（fringed）总的来说，这是一双简约、白色的牛仔短裤，具有线条图案、阔腿设计和高腰裤口。<|eot_id|>

d|>

You are a helpful assistant, 请用简体中文回答.<|eot_id|><|start_header_id|>user<|end_header_id|>

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<|eot_id|><|start_header_id|>assistant<|end_header_id|>

您所描述的裤子是：

  • 材质：牛仔布（denim）
  • 颜色：白色（white）
  • 风格：简约（simple）
  • 图案：线条（striped）
  • 裤长：短裤（short）
  • 裤型：阔腿裤（wide-leg）
  • 裤腰型：高腰（high-waisted）
  • 裤口：毛边（fringed）

总的来说，这是一双简约、白色的牛仔短裤，具有线条图案、阔腿设计和高腰裤口。<|eot_id|>

这篇关于LLaMA3（Meta）微调SFT实战Meta-Llama-3-8B-Instruct的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！


http://www.chinasem.cn/article/949067。

相关文章

网页解析 lxml 库--实战

网页解析 lxml 库--实战

lxml库使用流程 lxml 是 Python 的第三方解析库，完全使用 Python 语言编写，它对 XPath表达式提供了良好的支 持，因此能够了高效地解析 HTML/XML 文档。本节讲解如何通过 lxml 库解析 HTML 文档。 pip install lxml lxm| 库提供了一个 etree 模块，该模块专门用来解析 HTML/XML 文档，下面来介绍一下 lxml 库
阅读更多...
性能分析之MySQL索引实战案例

性能分析之MySQL索引实战案例

文章目录 一、前言二、准备三、MySQL索引优化四、MySQL 索引知识回顾五、总结 一、前言 在上一讲性能工具之 JProfiler 简单登录案例分析实战中已经发现SQL没有建立索引问题，本文将一起从代码层去分析为什么没有建立索引？ 开源ERP项目地址：https://gitee.com/jishenghua/JSH_ERP 二、准备 打开IDEA找到登录请求资源路径位置
阅读更多...
C#实战|大乐透选号器[6]：实现实时显示已选择的红蓝球数量

C#实战|大乐透选号器[6]：实现实时显示已选择的红蓝球数量

哈喽，你好啊，我是雷工。 关于大乐透选号器在前面已经记录了5篇笔记，这是第6篇； 接下来实现实时显示当前选中红球数量，蓝球数量； 以下为练习笔记。 01 效果演示 当选择和取消选择红球或蓝球时，在对应的位置显示实时已选择的红球、蓝球的数量； 02 标签名称 分别设置Label标签名称为：lblRedCount、lblBlueCount
阅读更多...
AI Toolkit + H100 GPU，一小时内微调最新热门文生图模型 FLUX

AI Toolkit + H100 GPU，一小时内微调最新热门文生图模型 FLUX

上个月，FLUX 席卷了互联网，这并非没有原因。他们声称优于 DALLE 3、Ideogram 和 Stable Diffusion 3 等模型，而这一点已被证明是有依据的。随着越来越多的流行图像生成工具（如 Stable Diffusion Web UI Forge 和 ComyUI）开始支持这些模型，FLUX 在 Stable Diffusion 领域的扩展将会持续下去。 自 FLU
阅读更多...
滚雪球学Java(87)：Java事务处理：JDBC的ACID属性与实战技巧！真有两下子！

滚雪球学Java(87)：Java事务处理：JDBC的ACID属性与实战技巧！真有两下子！

咦咦咦，各位小可爱，我是你们的好伙伴——bug菌，今天又来给大家普及Java SE啦，别躲起来啊，听我讲干货还不快点赞，赞多了我就有动力讲得更嗨啦！所以呀，养成先点赞后阅读的好习惯，别被干货淹没了哦~ 🏆本文收录于「滚雪球学Java」专栏，专业攻坚指数级提升，助你一臂之力，带你早日登顶🚀，欢迎大家关注&&收藏！持续更新中，up！up！up！！ 环境说明：Windows 10
阅读更多...
springboot实战学习（1）(开发模式与环境)

springboot实战学习（1）(开发模式与环境)

目录 一、实战学习的引言 （1）前后端的大致学习模块 （2）后端 （3）前端 二、开发模式 一、实战学习的引言 （1）前后端的大致学习模块 （2）后端 Validation：做参数校验Mybatis：做数据库的操作Redis：做缓存Junit：单元测试项目部署：springboot项目部署相关的知识 （3）前端 Vite：Vue项目的脚手架Router：路由Pina：状态管理Eleme
阅读更多...
深度学习实战：如何利用CNN实现人脸识别考勤系统

深度学习实战：如何利用CNN实现人脸识别考勤系统

1. 何为CNN及其在人脸识别中的应用 卷积神经网络（CNN）是深度学习中的核心技术之一，擅长处理图像数据。CNN通过卷积层提取图像的局部特征，在人脸识别领域尤其适用。CNN的多个层次可以逐步提取面部的特征，最终实现精确的身份识别。对于考勤系统而言，CNN可以自动从摄像头捕捉的视频流中检测并识别出员工的面部。 我们在该项目中采用了 RetinaFace 模型，它基于CNN的结构实现高效、精准的
阅读更多...
项目实战系列三: 家居购项目 第四部分

项目实战系列三: 家居购项目 第四部分

购物车 🌳购物车🍆显示购物车🍆更改商品数量🍆清空购物车&&删除商品 🌳生成订单 🌳购物车 需求分析 1.会员登陆后, 可以添加家居到购物车 2.完成购物车的设计和实现 3.每添加一个家居,购物车的数量+1, 并显示 程序框架图 1.新建src/com/zzw/furns/entity/CartItem.java, CartItem-家居项模型 /***
阅读更多...
Birt报表开发实战

Birt报表开发实战

我就截图描述得了，没什么含金量，看图基本明白的。 1.开始 a.创建报表文件 b.数据源配置 c.配置数据集 2.网格报表 拖拉式操作，很方便 3.预览效果 其他报表的操作也基本不难，就不扯了！ 2.级联参数 官方视频教程：http://demo.actuate.com/demos/cascade/cascade.html
阅读更多...
[yolov5] --- yolov5入门实战「土堆视频」

[yolov5] --- yolov5入门实战「土堆视频」

1 项目介绍及环境配置 下载yolov5 tags 5.0源码，https://github.com/ultralytics/yolov5/tree/v5.0，解压 Pycharm 中创建conda虚拟环境 激活conda虚拟环境 根据作者提供的requirements.txt文件，pip install -r requirements.txt 如果作者没有提供requirement.txt文件
阅读更多...

Copyright China编程 23002807@qq.com

沪ICP备2023033072号-2