使用 LLaMA Factory 微调 Llama-3 中文对话模型

已于 2024-04-22 22:30:34 修改
阅读量1.9k 收藏 3
点赞数 2
分类专栏： prompt aigc 大模型 文章标签： llama transformer AIGC
于 2024-04-22 21:21:46 首次发布
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.youkuaiyun.com/javastart/article/details/138092873
版权

原文：https://colab.research.google.com/drive/1d5KQtbemerlSDSxZIfAaWXhKr30QypiK?usp=sharing#scrollTo=gf60HoT633NY

请申请一个免费 T4 GPU 来运行该脚本

详细讲上面连接。需要科学上网

微调过程大约需要 50 分钟。

训练脚本：

from llmtuner import run_exp

%cd /content/LLaMA-Factory/

run_exp(dict(

  stage="sft",

  do_train=True,

  model_name_or_path="unsloth/llama-3-8b-Instruct-bnb-4bit",

  dataset="identity,alpaca_gpt4_en,alpaca_gpt4_zh",

  template="llama3",

  finetuning_type="lora",

  lora_target="all",

  output_dir="llama3_lora",

  per_device_train_batch_size=2,

  gradient_accumulation_steps=4,

  lr_scheduler_type="cosine",

  logging_steps=10,

  warmup_ratio=0.1,

  save_steps=1000,

  learning_rate=5e-5,

  num_train_epochs=3.0,

  max_samples=500,

  max_grad_norm=1.0,

  quantization_bit=4,

  loraplus_lr_ratio=16.0,

  use_unsloth=True,

  fp16=True,

))

训练过程日志

04/22/2024 04:10:40 - WARNING - llmtuner.hparams.parser - We recommend enable `upcast_layernorm` in quantized training.

WARNING:llmtuner.hparams.parser:We recommend enable `upcast_layernorm` in quantized training.

04/22/2024 04:10:40 - INFO - llmtuner.hparams.parser - Process rank: 0, device: cuda:0, n_gpu: 1, distributed training: False, compute dtype: torch.float16

INFO:llmtuner.hparams.parser:Process rank: 0, device: cuda:0, n_gpu: 1, distributed training: False, compute dtype: torch.float16
/usr/local/lib/python3.10/dist-packages/huggingface_hub/utils/_token.py:89: UserWarning: 
The secret `HF_TOKEN` does not exist in your Colab secrets.
To authenticate with the Hugging Face Hub, create a token in your settings tab (https://huggingface.co/settings/tokens), set it as secret in your Google Colab and restart your session.
You will be able to reuse this secret in all of your notebooks.
Please note that authentication is recommended but still optional to access public models or datasets.
  warnings.warn(
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:10:41,979 >> loading file tokenizer.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:10:41,980 >> loading file added_tokens.json from cache at None
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:10:41,982 >> loading file special_tokens_map.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/special_tokens_map.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:10:41,984 >> loading file tokenizer_config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer_config.json
[WARNING|logging.py:314] 2024-04-22 04:10:42,384 >> Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.

04/22/2024 04:10:42 - INFO - llmtuner.data.template - Replace eos token: <|eot_id|>

INFO:llmtuner.data.template:Replace eos token: <|eot_id|>

04/22/2024 04:10:42 - INFO - llmtuner.data.loader - Loading dataset identity.json...

INFO:llmtuner.data.loader:Loading dataset identity.json...

04/22/2024 04:10:42 - WARNING - llmtuner.data.utils - Checksum failed: mismatched SHA-1 hash value at data/identity.json.

WARNING:llmtuner.data.utils:Checksum failed: mismatched SHA-1 hash value at data/identity.json.

Generating train split: 

 91/0 [00:00<00:00, 1640.44 examples/s]

Converting format of dataset: 100%

 91/91 [00:00<00:00, 2822.67 examples/s]

04/22/2024 04:10:42 - INFO - llmtuner.data.loader - Loading dataset alpaca_gpt4_data_en.json...

INFO:llmtuner.data.loader:Loading dataset alpaca_gpt4_data_en.json...

Generating train split: 

 52002/0 [00:00<00:00, 117346.95 examples/s]

Converting format of dataset: 100%

 500/500 [00:00<00:00, 14816.36 examples/s]

04/22/2024 04:10:43 - INFO - llmtuner.data.loader - Loading dataset alpaca_gpt4_data_zh.json...

INFO:llmtuner.data.loader:Loading dataset alpaca_gpt4_data_zh.json...

Generating train split: 

 48818/0 [00:00<00:00, 91511.83 examples/s]

Converting format of dataset: 100%

 500/500 [00:00<00:00, 11785.79 examples/s]

Running tokenizer on dataset: 100%

 1091/1091 [00:00<00:00, 1358.62 examples/s]

[INFO|configuration_utils.py:728] 2024-04-22 04:10:45,417 >> loading configuration file config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/config.json
[INFO|configuration_utils.py:791] 2024-04-22 04:10:45,419 >> Model config LlamaConfig {
  "_name_or_path": "unsloth/llama-3-8b-Instruct-bnb-4bit",
  "architectures": [
    "LlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 128000,
  "eos_token_id": 128001,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 14336,
  "max_position_embeddings": 8192,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 8,
  "pretraining_tp": 1,
  "quantization_config": {
    "_load_in_4bit": true,
    "_load_in_8bit": false,
    "bnb_4bit_compute_dtype": "bfloat16",
    "bnb_4bit_quant_type": "nf4",
    "bnb_4bit_use_double_quant": true,
    "llm_int8_enable_fp32_cpu_offload": false,
    "llm_int8_has_fp16_weight": false,
    "llm_int8_skip_modules": null,
    "llm_int8_threshold": 6.0,
    "load_in_4bit": true,
    "load_in_8bit": false,
    "quant_method": "bitsandbytes"
  },
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 500000.0,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.38.2",
  "use_cache": true,
  "vocab_size": 128256
}


input_ids:
[128000, 128006, 9125, 128007, 271, 2675, 527, 264, 11190, 18328, 13, 128009, 128006, 882, 128007, 271, 6151, 128009, 128006, 78191, 128007, 271, 9906, 0, 358, 1097, 445, 81101, 30653, 7496, 11, 459, 15592, 18328, 8040, 555, 445, 8921, 4940, 17367, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]
inputs:
<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a helpful assistant.<|eot_id|><|start_header_id|>user<|end_header_id|>

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

Hello! I am Llama-Chinese, an AI assistant developed by LLaMA Factory. How can I assist you today?<|eot_id|>
label_ids:
[-100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, 9906, 0, 358, 1097, 445, 81101, 30653, 7496, 11, 459, 15592, 18328, 8040, 555, 445, 8921, 4940, 17367, 13, 2650, 649, 358, 7945, 499, 3432, 30, 128009]
labels:
Hello! I am Llama-Chinese, an AI assistant developed by LLaMA Factory. How can I assist you today?<|eot_id|>
04/22/2024 04:10:45 - INFO - llmtuner.model.patcher - Loading ?-bit BITSANDBYTES-quantized model.

INFO:llmtuner.model.patcher:Loading ?-bit BITSANDBYTES-quantized model.
[INFO|configuration_utils.py:728] 2024-04-22 04:10:45,579 >> loading configuration file config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/config.json
[INFO|configuration_utils.py:791] 2024-04-22 04:10:45,581 >> Model config LlamaConfig {
  "_name_or_path": "unsloth/llama-3-8b-Instruct-bnb-4bit",
  "architectures": [
    "LlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 128000,
  "eos_token_id": 128001,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 14336,
  "max_position_embeddings": 8192,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 8,
  "pretraining_tp": 1,
  "quantization_config": {
    "_load_in_4bit": true,
    "_load_in_8bit": false,
    "bnb_4bit_compute_dtype": "bfloat16",
    "bnb_4bit_quant_type": "nf4",
    "bnb_4bit_use_double_quant": true,
    "llm_int8_enable_fp32_cpu_offload": false,
    "llm_int8_has_fp16_weight": false,
    "llm_int8_skip_modules": null,
    "llm_int8_threshold": 6.0,
    "load_in_4bit": true,
    "load_in_8bit": false,
    "quant_method": "bitsandbytes"
  },
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 500000.0,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.38.2",
  "use_cache": true,
  "vocab_size": 128256
}

[INFO|configuration_utils.py:728] 2024-04-22 04:10:45,634 >> loading configuration file config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/config.json
[INFO|configuration_utils.py:791] 2024-04-22 04:10:45,636 >> Model config LlamaConfig {
  "_name_or_path": "unsloth/llama-3-8b-Instruct-bnb-4bit",
  "architectures": [
    "LlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 128000,
  "eos_token_id": 128001,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 14336,
  "max_position_embeddings": 8192,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 8,
  "pretraining_tp": 1,
  "quantization_config": {
    "_load_in_4bit": true,
    "_load_in_8bit": false,
    "bnb_4bit_compute_dtype": "bfloat16",
    "bnb_4bit_quant_type": "nf4",
    "bnb_4bit_use_double_quant": true,
    "llm_int8_enable_fp32_cpu_offload": false,
    "llm_int8_has_fp16_weight": false,
    "llm_int8_skip_modules": null,
    "llm_int8_threshold": 6.0,
    "load_in_4bit": true,
    "load_in_8bit": false,
    "quant_method": "bitsandbytes"
  },
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 500000.0,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.38.2",
  "use_cache": true,
  "vocab_size": 128256
}

[INFO|configuration_utils.py:728] 2024-04-22 04:10:45,702 >> loading configuration file config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/config.json
[INFO|configuration_utils.py:791] 2024-04-22 04:10:45,704 >> Model config LlamaConfig {
  "_name_or_path": "unsloth/llama-3-8b-Instruct-bnb-4bit",
  "architectures": [
    "LlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 128000,
  "eos_token_id": 128001,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 14336,
  "max_position_embeddings": 8192,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 8,
  "pretraining_tp": 1,
  "quantization_config": {
    "_load_in_4bit": true,
    "_load_in_8bit": false,
    "bnb_4bit_compute_dtype": "bfloat16",
    "bnb_4bit_quant_type": "nf4",
    "bnb_4bit_use_double_quant": true,
    "llm_int8_enable_fp32_cpu_offload": false,
    "llm_int8_has_fp16_weight": false,
    "llm_int8_skip_modules": null,
    "llm_int8_threshold": 6.0,
    "load_in_4bit": true,
    "load_in_8bit": false,
    "quant_method": "bitsandbytes"
  },
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 500000.0,
  "tie_word_embeddings": false,
  "torch_dtype": "float16",
  "transformers_version": "4.38.2",
  "use_cache": true,
  "vocab_size": 128256
}


==((====))==  Unsloth: Fast Llama patching release 2024.4
   \\   /|    GPU: Tesla T4. Max memory: 14.748 GB. Platform = Linux.
O^O/ \_/ \    Pytorch: 2.2.1+cu121. CUDA = 7.5. CUDA Toolkit = 12.1.
\        /    Bfloat16 = FALSE. Xformers = 0.0.25.post1. FA = False.
 "-____-"     Free Apache license: GitHub - unslothai/unsloth: Finetune Llama 3, Mistral & Gemma LLMs 2-5x faster with 80% less memory

[INFO|modeling_utils.py:3257] 2024-04-22 04:10:45,813 >> loading weights file model.safetensors from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/model.safetensors
[INFO|modeling_utils.py:1400] 2024-04-22 04:10:45,863 >> Instantiating LlamaForCausalLM model under default dtype torch.float16.
[INFO|configuration_utils.py:845] 2024-04-22 04:10:45,871 >> Generate config GenerationConfig {
  "bos_token_id": 128000,
  "eos_token_id": 128001
}

[INFO|modeling_utils.py:3992] 2024-04-22 04:11:13,469 >> All model checkpoint weights were used when initializing LlamaForCausalLM.

[INFO|modeling_utils.py:4000] 2024-04-22 04:11:13,472 >> All the weights of LlamaForCausalLM were initialized from the model checkpoint at unsloth/llama-3-8b-Instruct-bnb-4bit.
If your task is similar to the task the model of the checkpoint was trained on, you can already use LlamaForCausalLM for predictions without further training.
[INFO|configuration_utils.py:800] 2024-04-22 04:11:13,539 >> loading configuration file generation_config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/generation_config.json
[INFO|configuration_utils.py:845] 2024-04-22 04:11:13,540 >> Generate config GenerationConfig {
  "bos_token_id": 128000,
  "eos_token_id": 128001
}

tokenizer_config.json: 100%

 51.0k/51.0k [00:00<00:00, 2.14MB/s]

tokenizer.json: 100%

 9.08M/9.08M [00:00<00:00, 60.7MB/s]

special_tokens_map.json: 100%

 449/449 [00:00<00:00, 31.3kB/s]

[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,466 >> loading file tokenizer.json from cache at huggingface_tokenizers_cache/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,468 >> loading file added_tokens.json from cache at None
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,469 >> loading file special_tokens_map.json from cache at huggingface_tokenizers_cache/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/special_tokens_map.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,472 >> loading file tokenizer_config.json from cache at huggingface_tokenizers_cache/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer_config.json
[WARNING|logging.py:314] 2024-04-22 04:11:14,881 >> Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,935 >> loading file tokenizer.json from cache at huggingface_tokenizers_cache/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,936 >> loading file added_tokens.json from cache at None
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,937 >> loading file special_tokens_map.json from cache at huggingface_tokenizers_cache/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/special_tokens_map.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 04:11:14,939 >> loading file tokenizer_config.json from cache at huggingface_tokenizers_cache/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer_config.json
[WARNING|logging.py:314] 2024-04-22 04:11:15,312 >> Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.

04/22/2024 04:11:16 - INFO - llmtuner.model.patcher - Gradient checkpointing enabled.

INFO:llmtuner.model.patcher:Gradient checkpointing enabled.

04/22/2024 04:11:16 - INFO - llmtuner.model.adapter - Fine-tuning method: LoRA

INFO:llmtuner.model.adapter:Fine-tuning method: LoRA

04/22/2024 04:11:16 - INFO - llmtuner.model.utils - Found linear modules: k_proj,o_proj,down_proj,v_proj,up_proj,q_proj,gate_proj

INFO:llmtuner.model.utils:Found linear modules: k_proj,o_proj,down_proj,v_proj,up_proj,q_proj,gate_proj
[WARNING|logging.py:329] 2024-04-22 04:11:16,731 >> Unsloth 2024.4 patched 32 layers with 32 QKV layers, 32 O layers and 32 MLP layers.

04/22/2024 04:11:16 - INFO - llmtuner.model.loader - trainable params: 20971520 || all params: 8051232768 || trainable%: 0.2605

INFO:llmtuner.model.loader:trainable params: 20971520 || all params: 8051232768 || trainable%: 0.2605
[INFO|trainer.py:601] 2024-04-22 04:11:16,796 >> Using auto half precision backend

04/22/2024 04:11:17 - INFO - llmtuner.train.utils - Using LoRA+ optimizer with loraplus lr ratio 16.00.

INFO:llmtuner.train.utils:Using LoRA+ optimizer with loraplus lr ratio 16.00.
[WARNING|logging.py:329] 2024-04-22 04:11:17,203 >> ==((====))==  Unsloth - 2x faster free finetuning | Num GPUs = 1
   \\   /|    Num examples = 1,091 | Num Epochs = 3
O^O/ \_/ \    Batch size per device = 2 | Gradient Accumulation steps = 4
\        /    Total batch size = 8 | Total steps = 408
 "-____-"     Number of trainable parameters = 20,971,520

 [408/408 48:57, Epoch 2/3]

StepTraining Loss
101.568300
201.478600
301.298700
401.188600
501.185700
601.200300
701.249100
801.213600
901.255900
1001.186000
1101.210600
1201.216200
1301.111400
1401.077700
1500.906100
1600.895100
1700.981500
1800.759400
1900.834800
2000.816900
2100.773200
2200.946500
2300.764600
2400.914700
2500.864800
2600.840600
2700.853600
2800.745800
2900.500800
3000.597600
3100.616400
3200.574100
3300.490300
3400.602800
3500.563700
3600.552900
3700.574400
3800.468200
3900.549200
4000.528500

[INFO|<string>:460] 2024-04-22 05:00:27,815 >> 

Training completed. Do not forget to share your model on huggingface.co/models =)


[INFO|trainer.py:3067] 2024-04-22 05:00:27,822 >> Saving model checkpoint to llama3_lora
[INFO|configuration_utils.py:728] 2024-04-22 05:00:28,263 >> loading configuration file config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/config.json
[INFO|configuration_utils.py:791] 2024-04-22 05:00:28,266 >> Model config LlamaConfig {
  "_name_or_path": "meta-llama/Meta-Llama-3-8B-Instruct",
  "architectures": [
    "LlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 128000,
  "eos_token_id": 128001,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 14336,
  "max_position_embeddings": 8192,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 8,
  "pretraining_tp": 1,
  "quantization_config": {
    "_load_in_4bit": true,
    "_load_in_8bit": false,
    "bnb_4bit_compute_dtype": "bfloat16",
    "bnb_4bit_quant_type": "nf4",
    "bnb_4bit_use_double_quant": true,
    "llm_int8_enable_fp32_cpu_offload": false,
    "llm_int8_has_fp16_weight": false,
    "llm_int8_skip_modules": null,
    "llm_int8_threshold": 6.0,
    "load_in_4bit": true,
    "load_in_8bit": false,
    "quant_method": "bitsandbytes"
  },
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 500000.0,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.38.2",
  "use_cache": true,
  "vocab_size": 128256
}

[INFO|tokenization_utils_base.py:2459] 2024-04-22 05:00:28,538 >> tokenizer config file saved in llama3_lora/tokenizer_config.json
[INFO|tokenization_utils_base.py:2468] 2024-04-22 05:00:28,541 >> Special tokens file saved in llama3_lora/special_tokens_map.json
[INFO|modelcard.py:450] 2024-04-22 05:00:28,827 >> Dropping the following result as it does not have all the necessary fields:
{'task': {'name': 'Causal Language Modeling', 'type': 'text-generation'}}

***** train metrics *****
  epoch                    =       2.99
  total_flos               = 32079633GF
  train_loss               =     0.8929
  train_runtime            = 0:49:10.61
  train_samples_per_second =      1.109
  train_steps_per_second   =      0.138

推理：

from llmtuner import ChatModel

from llmtuner.extras.misc import torch_gc

%cd /content/LLaMA-Factory/

chat_model = ChatModel(dict(

  model_name_or_path="unsloth/llama-3-8b-Instruct-bnb-4bit",

  adapter_name_or_path="llama3_lora",

  finetuning_type="lora",

  template="llama3",

))

messages = []

while True:

  query = input("\nUser: ")

  if query.strip() == "exit":

    torch_gc()

    break

  if query.strip() == "clear":

    messages = []

    torch_gc()

    print("History has been removed.")

    continue

  messages.append({"role": "user", "content": query})

  print("Assistant: ", end="", flush=True)

  response = ""

  for new_text in chat_model.stream_chat(messages):

    print(new_text, end="", flush=True)

    response += new_text

  print()

  messages.append({"role": "assistant", "content": response})

推理执行日志

/usr/local/lib/python3.10/dist-packages/huggingface_hub/utils/_token.py:89: UserWarning: 
The secret `HF_TOKEN` does not exist in your Colab secrets.
To authenticate with the Hugging Face Hub, create a token in your settings tab (https://huggingface.co/settings/tokens), set it as secret in your Google Colab and restart your session.
You will be able to reuse this secret in all of your notebooks.
Please note that authentication is recommended but still optional to access public models or datasets.
  warnings.warn(
[INFO|tokenization_utils_base.py:2046] 2024-04-22 05:12:13,951 >> loading file tokenizer.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 05:12:13,953 >> loading file added_tokens.json from cache at None
[INFO|tokenization_utils_base.py:2046] 2024-04-22 05:12:13,957 >> loading file special_tokens_map.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/special_tokens_map.json
[INFO|tokenization_utils_base.py:2046] 2024-04-22 05:12:13,959 >> loading file tokenizer_config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/tokenizer_config.json
[WARNING|logging.py:314] 2024-04-22 05:12:14,407 >> Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.

04/22/2024 05:12:14 - INFO - llmtuner.data.template - Replace eos token: <|eot_id|>

INFO:llmtuner.data.template:Replace eos token: <|eot_id|>
[INFO|configuration_utils.py:728] 2024-04-22 05:12:14,462 >> loading configuration file config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/config.json
[INFO|configuration_utils.py:791] 2024-04-22 05:12:14,464 >> Model config LlamaConfig {
  "_name_or_path": "unsloth/llama-3-8b-Instruct-bnb-4bit",
  "architectures": [
    "LlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 128000,
  "eos_token_id": 128001,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 14336,
  "max_position_embeddings": 8192,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 8,
  "pretraining_tp": 1,
  "quantization_config": {
    "_load_in_4bit": true,
    "_load_in_8bit": false,
    "bnb_4bit_compute_dtype": "bfloat16",
    "bnb_4bit_quant_type": "nf4",
    "bnb_4bit_use_double_quant": true,
    "llm_int8_enable_fp32_cpu_offload": false,
    "llm_int8_has_fp16_weight": false,
    "llm_int8_skip_modules": null,
    "llm_int8_threshold": 6.0,
    "load_in_4bit": true,
    "load_in_8bit": false,
    "quant_method": "bitsandbytes"
  },
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 500000.0,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.38.2",
  "use_cache": true,
  "vocab_size": 128256
}


04/22/2024 05:12:14 - INFO - llmtuner.model.patcher - Loading ?-bit BITSANDBYTES-quantized model.

INFO:llmtuner.model.patcher:Loading ?-bit BITSANDBYTES-quantized model.

04/22/2024 05:12:14 - INFO - llmtuner.model.patcher - Using KV cache for faster generation.

INFO:llmtuner.model.patcher:Using KV cache for faster generation.
[INFO|modeling_utils.py:3257] 2024-04-22 05:12:14,509 >> loading weights file model.safetensors from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/model.safetensors
[INFO|modeling_utils.py:1400] 2024-04-22 05:12:14,560 >> Instantiating LlamaForCausalLM model under default dtype torch.float16.
[INFO|configuration_utils.py:845] 2024-04-22 05:12:14,569 >> Generate config GenerationConfig {
  "bos_token_id": 128000,
  "eos_token_id": 128001
}

[INFO|modeling_utils.py:3992] 2024-04-22 05:12:21,290 >> All model checkpoint weights were used when initializing LlamaForCausalLM.

[INFO|modeling_utils.py:4000] 2024-04-22 05:12:21,293 >> All the weights of LlamaForCausalLM were initialized from the model checkpoint at unsloth/llama-3-8b-Instruct-bnb-4bit.
If your task is similar to the task the model of the checkpoint was trained on, you can already use LlamaForCausalLM for predictions without further training.
[INFO|configuration_utils.py:800] 2024-04-22 05:12:21,357 >> loading configuration file generation_config.json from cache at /root/.cache/huggingface/hub/models--unsloth--llama-3-8b-Instruct-bnb-4bit/snapshots/3fe39f383d040c52882f7f9faaf600dc57c01a8d/generation_config.json
[INFO|configuration_utils.py:845] 2024-04-22 05:12:21,359 >> Generate config GenerationConfig {
  "bos_token_id": 128000,
  "eos_token_id": 128001
}


04/22/2024 05:12:21 - INFO - llmtuner.model.adapter - Fine-tuning method: LoRA

INFO:llmtuner.model.adapter:Fine-tuning method: LoRA

04/22/2024 05:12:22 - INFO - llmtuner.model.adapter - Loaded adapter(s): llama3_lora

INFO:llmtuner.model.adapter:Loaded adapter(s): llama3_lora

04/22/2024 05:12:22 - INFO - llmtuner.model.loader - all params: 8051232768

INFO:llmtuner.model.loader:all params: 8051232768

User: 你是谁
Assistant: 您好，我是 Llama-Chinese，一个由 LLaMA Factory 开发的人工智能助手。我可以根据用户的需求提供相应的回答和帮助。

User: 我的蓝牙耳机坏了，我该去看牙科还是耳鼻喉科？
Assistant: 很抱歉听到您的小配件出了问题。蓝牙耳机的问题与牙科或耳鼻喉科无关。您可以尝试与制造商或当地的维修店联系，进行维修或更换。

User: 谢谢你的帮助
Assistant: 不客气。我是人工智能助手 Llama-Chinese，很高兴能帮到您。

User: exit

LLaMA Factory单机微调的实战教程
herosunly的博客
04-24 19万+
本文主要介绍了使用LLaMA Factory对大模型进行单机微调的实战教程，希望对学习大语言模型的同学们有所帮助。 文章目录 1. 前言 2. 配置环境 2.1 安装虚拟环境 2.2 安装依赖库 2.3 模型选型与下载 3. 运行代码 3.1 构建训练数据 3.2 修改配置文件 3.3 微调
开源模型应用落地-qwen2-7b-instruct-LoRA微调-LLaMA-Factory-单机单卡-V100（八）
没有卑微的工作，只有卑微的心态，与其抱怨，不如埋头实干
08-21 8192
使用LLaMA-Factory高效微调qwen2-7b-instruct
用 LLama-Factory 训练和微调 LLama3，打造你的专属 AI 模型！
数智笔记
05-19 2404
微调模型涉及调整预训练或基础模型的参数，使其能够用于特定任务或数据集，从而提升其性能和准确性。这个过程包括为模型提供新的数据，并修改其权重、偏差和某些参数以最小化损失和成本。通过这样做，这个新模型可以在任何新任务或数据集上表现良好，而无需从头开始，从而节省时间和资源。通常，当一个新的大型语言模型（LLM）创建时，它会在一个大型文本数据语料库上进行训练，其中可能包含潜在有害或不良内容。在预训练或初始训练阶段之后，模型会被微调加入安全措施，确保其避免生成有害或不良的响应。然而，这种方法并非完美无缺。
教程：利用LLaMA_Factory微调llama3:8b大模型
m0_60683691的博客
05-06 4249
在存放模型空白处右键，点击git bash here，下载llama3模型至本地由于文件较大，因此该过程比较久，保持下载窗口打开，知道出现下一命令行表示下载完成。
【大模型系列篇】初探Unsloth微调DeepSeek-R1蒸馏模型
最新发布
木亦汐丫
02-27 1848
🔥🔥🔥 本文初探使用Unsloth微调DeepSeek-R1蒸馏模型DeepSeek-R1-Distill-Qwen-7B，使用了COT思维链数据进行Lora微调，微调后模型合并权重导出，并使用llama.cpp进行gguf格式转换和量化，通过ollama创建并加载模型，进行部署。
Llama-Factory + Ollama 打造属于自己的中文版 Llama3
hyang1974的博客
05-05 8996
Meta Llama3发布已经有一小段时间了。和以往的原始 Llama 模型一样，Llama 3 对中文的支持效果欠佳。本文则探讨如何使用 Llama-Factory 对 Llama3 进行中文微调的具体过程，并通过 Ollama 本地部署中文微调的 Llama3 模型，打造属于自己的个性化的 Llama3 LLM 。
使用LLaMA-Factory微调Llama3大模型
T.S.Hao的博客
06-19 3085
本文详细介绍了如何在本地基于llama3中文预训练模型，使用LLaMA-Factory微调模型，实现特定领域的下游任务。
10 分钟，教你如何用 LLama-Factory 训练和微调 LLama3 模型
了解关于云主机、GPU、AI、数据库托管、Kubernetes等相关技术知识，以及DigitalOcean云平台使用教程
05-21 9840
在本文中，我们将了解如何使用 LLama Index 对 LLama3 进行微调。更重要的是，你只需要通过非常简单的步骤和几行代码，就能实现这一点。
LLaMA-Factory+qwen多轮对话微调测评
kk的博客
04-02 8979
LLaMA-Factory+qwen多轮对话微调
LLaMA Factory多卡微调的实战教程
热门推荐
herosunly的博客
04-25 19万+
本文主要介绍了LLaMA Factory多卡微调的实战教程，希望对学习大语言模型的同学们有所帮助。 文章目录 1. 前言 2. 配置环境 2.1 安装虚拟环境 2.2 安装依赖库 2.3 模型选型与下载 2.4 配置多卡环境 3. 运行代码 3.1 构建训练数据 3.2 修改配置文件 3.3 微调 3.3.1 zero3微调 3.3.2 启动API 3.3.3 模型测试
LLM - 使用 LLaMA-Factory 微调 Qwen2-VL SFT(LoRA) 图像数据集 教程 (2)
AGI
11-12 2636
LLaMA-Factory 是开源的大模型微调框架，用于高效地微调和部署大语言模型，支持多种预训练模型和微调算法，提供完整的工具和接口，对于预训练的模型进行定制化的训练和调整，以适应特定的应用场景。
基于LLaMA-Factory微调Llama3
zhishi0000的博客
05-29 1692
本文简要介绍下基于LLaMA-Factory的llama3 8B模型的微调过程升级到24.0版本创建微调脚本Step 6. 测试微调效果​。
AI大模型企业应用实战：Prompt让LLM理解知识
2401_84204413的博客
06-25 2464
• 能够完成时下热门大模型垂直领域模型训练能力，提高程序员的编码能力： 大模型应用开发需要掌握机器学习算法、深度学习框架等技术，这些技术的掌握可以提高程序员的编码能力和分析能力，让程序员更加熟练地编写高质量的代码。• 基于大模型和企业数据AI应用开发，实现大模型理论、掌握GPU算力、硬件、LangChain开发框架和项目实战技能， 学会Fine-tuning垂直训练大模型（数据准备、数据蒸馏、大模型部署）一站式掌握；第五阶段： 大模型微调开发借助以大健康、新零售、新媒体领域构建适合当前领域大模型；
大模型微调之 使用 LLaMA-Factory 微调 Llama3
段智华的博客
04-25 3322
大模型微调之 使用 LLaMA-Factory 微调 Llama3使用 LLaMA Factory 微调 Llama-3 中文对话模型 请申请一个免费 T4 GPU 来运行该脚本 运行结果为： 检查 GPU 环境 更新自我认知数据集 可以自由修改 NAME 和 AUTHOR 变量的内容。 identity.json的文件内容为： 同时，加载alpaca_gpt4_en,alpaca_gpt4_zh数据集，对应alpaca_gpt4_data_en.json、alpaca_gpt4_data_zh.jso
llama3 微调教程之 llama factory 的 安装部署与模型微调过程，模型量化和gguf转换。
失业求捞！
05-17 1万+
本文记录了从环境部署到微调模型、效果测试的全过程，以及遇到几个常见问题的解决办法，亲测可用（The installed version of bitsandbytes was compiled without GPU support. NotImplementedError: Architecture ‘LlamaForCausalLM’ not supported!
LlamaFactory 进行大模型 llama3 微调，轻松上手体验学习
2401_84033492的博客
04-24 2111
节前，我们星球组织了一场算法岗技术&面试讨论会，邀请了一些互联网大厂朋友、参加社招和校招面试的同学，针对算法岗技术趋势、大模型落地项目经验分享、新手如何入门算法岗、该如何准备、面试常考点分享等热门话题进行了深入的讨论。
LLAMA3 微调与部署
t951984189的博客
05-10 718
是一个快速微调 Mistral，Gema，Llama 等模型的工具，速度提高 2-5 倍，内存减少 80%！安装过程需要耐心等待，大部份问题都是由于网络问题导致的，如果遇到问题，可以尝试多次安装，或者使用代理。建议在 Linux 环境下使用，否则可能你需要处理各种各样的奇怪的问题。要好的效果，你得认真准备你要训练的数据集，::: warning 注意。
微调神器LLaMA-Factory官方保姆级教程来了，从环境搭建到模型训练评估全覆盖
2401_85377976的博客
08-13 1519
数据集的格式要求在不同的阶段是不同的，本教程以sft阶段的数据集需求，将以系统自带的identity数据集和将自定义的一个商品文案生成数据集为例，介绍数据集的使用。更多详情可以在https://github.com/hiyouga/LLaMA-Factory/blob/main/data/README_zh.md 中找到相关解释。系统目前支持 alpaca 和sharegpt两种数据格式，以alpaca为例，整个数据集是一个json对象的list，具体数据格式为。
llama-factory微调DeepSeek-R1-Distill-Qwen-7B
02-27
### 使用 LLaMA-Factory 对 DeepSeek-R1-Distill-Qwen-7B 模型进行微调 对于希望使用 `LLaMA-Factory` 工具对特定模型如 `DeepSeek-R1-Distill-Qwen-7B` 进行微调的情况，可以遵循以下指导： #### 准备环境与资源 确保已经按照官方说明克隆了项目仓库并安装依赖项。这一步骤至关重要，因为后续操作均基于此开发环境中执行。 ```bash git clone http://developer.sourcefind.cn/codes/OpenDAS/llama-factory.git cd llama-factory && pip install -e ".[torch,metrics]" ``` 上述命令用于获取最新版本的源码以及设置必要的Python包支持[^1]。 #### 配置微调参数文件 创建或编辑配置YAML文件以定义具体的超参设定、数据路径及其他选项。针对不同任务需求调整这些参数能够显著影响最终效果。例如，在处理图像分类问题时，可能需要特别指定输入特征维度等细节。 假设有一个名为 `my_custom_finetune_config.yaml` 的配置文件，其中包含了关于目标领域（比如时尚物品识别）、训练集位置以及其他必要信息的内容描述。 #### 执行微调过程 利用预构建脚本启动实际的微调流程。这里假定采用的是低秩适应（LoRA）技术来进行高效迁移学习，则对应的CLI指令可能是这样的形式： ```bash llamafactory-cli train examples/train_lora/mytrain_lora_sft.yaml ``` 这条语句会读取之前准备好的配置文档，并据此开始优化给定的基础架构——即此处提到的 `DeepSeek-R1-Distill-Qwen-7B` ——使之更贴合于新的应用场景下的表现期望[^2]。 #### 测试改进后的性能 一旦完成一轮或多轮迭代更新之后，应当及时评估新版本的表现差异。借助内置Web界面功能可以直接加载最新的检查点权重，并通过交互方式快速验证某些具体实例上的变化趋势。 ```bash llamafactory-cli webui ``` 此时可以在图形界面上轻松上传待测样本（如图片），观察经过定制化改造过的网络结构能否更加精准地给出预期类别标签[^3]。 #### 应用场景中的考量 考虑到实际应用背景中存在将第三方标注体系转换成本地标准的需求，直接通过对基础AI系统的针对性再教育来减少中间环节误差不失为一种有效策略。这种方式不仅有助于提升整体判断精度，同时也简化了后期维护工作流的设计复杂度[^4]。
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值