<基于“通义千问1.8B基础模型”和“SWIFT微调框架”的情感分类模型>介绍
本项目使用 SWIFT (Scalable lightWeight Infrastructure for Fine-Tuning) 微调框架，选择采用LoRA微调方法，对 通义千问1.8B 基础模型进行微调，以实现对句子的情感分类任务。
项目来源： 2024年1月《阿里魔搭社区 AI大模型技术与应用实战训练营》(西安交通大学) 大作业一
戴锦程
实验环境
modelscope平台免费实例-GPU环境（8核 32GB 显存24G (A10显卡)；预装 ModelScope Library; 预装镜像 Ubuntu22.04 cuda12.1.0 python3.10 pytorch2.1.2 tensorflow2.14.0)
需手动安装微调框架库 swift 以及所需的其他第三方库，可以通过下方代码进行环境准备：
git clone https://github.com/modelscope/swift.git
cd swift
pip install -e .[llm]
训练方法
数据集
京东商品评论情感分析数据集 (jdsentimentzh)
微调的基础模型
通义千问1.8B (qwen18b)
微调方法
LoRA
其他超参数
sft参数 (sfc_args.json)
{
"model_type": "qwen-1_8b",
"model_id_or_path": "qwen/Qwen-1_8B",
"model_revision": "master",
"model_cache_dir": null,
"sft_type": "lora",
"freeze_parameters": 0.0,
"additional_trainable_parameters": [],
"tuner_backend": "swift",
"template_type": "default-generation",
"output_dir": "/mnt/workspace/output/qwen-1_8b/v0-20240128-024328",
"add_output_dir_suffix": true,
"ddp_backend": "nccl",
"seed": 42,
"resume_from_checkpoint": null,
"dtype": "bf16",
"dataset": [
"jd-sentiment-zh"
],
"dataset_seed": 42,
"dataset_test_ratio": 0.01,
"train_dataset_sample": 2000,
"val_dataset_sample": null,
"system": null,
"max_length": 2048,
"truncation_strategy": "delete",
"check_dataset_strategy": "none",
"custom_train_dataset_path": [],
"custom_val_dataset_path": [],
"self_cognition_sample": 0,
"model_name": [
null,
null
],
"model_author": [
null,
null
],
"quantization_bit": 0,
"bnb_4bit_comp_dtype": "bf16",
"bnb_4bit_quant_type": "nf4",
"bnb_4bit_use_double_quant": true,
"lora_target_modules": [
"c_attn"
],
"lora_rank": 8,
"lora_alpha": 32,
"lora_dropout_p": 0.05,
"lora_bias_trainable": "none",
"lora_modules_to_save": [],
"lora_dtype": "fp32",
"neftune_noise_alpha": null,
"gradient_checkpointing": true,
"deepspeed_config_path": null,
"batch_size": 1,
"eval_batch_size": 1,
"num_train_epochs": 1,
"max_steps": -1,
"optim": "adamw_torch",
"adam_beta1": 0.9,
"adam_beta2": 0.999,
"learning_rate": 0.0001,
"weight_decay": 0.01,
"gradient_accumulation_steps": 16,
"max_grad_norm": 0.5,
"predict_with_generate": false,
"lr_scheduler_type": "linear",
"warmup_ratio": 0.05,
"eval_steps": 50,
"save_steps": 50,
"save_only_model": false,
"save_total_limit": 2,
"logging_steps": 5,
"dataloader_num_workers": 1,
"push_to_hub": false,
"hub_model_id": "qwen-1_8b-lora",
"hub_private_repo": true,
"push_hub_strategy": "push_best",
"hub_token": null,
"test_oom_error": false,
"disable_tqdm": false,
"lazy_tokenize": false,
"preprocess_num_proc": 1,
"use_flash_attn": null,
"ignore_args_error": false,
"check_model_is_latest": true,
"logging_dir": "/mnt/workspace/output/qwen-1_8b/v0-20240128-024328/runs",
"report_to": [
"all"
],
"acc_strategy": "token",
"save_on_each_node": true,
"evaluation_strategy": "steps",
"save_strategy": "steps",
"save_safetensors": true,
"max_new_tokens": 2048,
"do_sample": true,
"temperature": 0.3,
"top_k": 20,
"top_p": 0.7,
"repetition_penalty": 1.0,
"num_beams": 1,
"per_device_train_batch_size": null,
"per_device_eval_batch_size": null,
"only_save_model": null,
"neftune_alpha": null,
"torch_dtype": "torch.bfloat16",
"fp16": false,
"bf16": true,
"bnb_4bit_compute_dtype": "torch.bfloat16",
"load_in_4bit": false,
"load_in_8bit": false,
"train_sampler_random": true,
"deepspeed": null
}
训练参数 (training_args.json)
{
"output_dir": "/mnt/workspace/output/qwen-1_8b/v0-20240128-024328",
"overwrite_output_dir": false,
"do_train": false,
"do_eval": true,
"do_predict": false,
"evaluation_strategy": "steps",
"prediction_loss_only": false,
"per_device_train_batch_size": 1,
"per_device_eval_batch_size": 1,
"per_gpu_train_batch_size": null,
"per_gpu_eval_batch_size": null,
"gradient_accumulation_steps": 16,
"eval_accumulation_steps": null,
"eval_delay": 0,
"learning_rate": 0.0001,
"weight_decay": 0.01,
"adam_beta1": 0.9,
"adam_beta2": 0.999,
"adam_epsilon": 1e-08,
"max_grad_norm": 0.5,
"num_train_epochs": 1,
"max_steps": -1,
"lr_scheduler_type": "linear",
"lr_scheduler_kwargs": {},
"warmup_ratio": 0.05,
"warmup_steps": 0,
"log_level": "passive",
"log_level_replica": "warning",
"log_on_each_node": true,
"logging_dir": "/mnt/workspace/output/qwen-1_8b/v0-20240128-024328/runs",
"logging_strategy": "steps",
"logging_first_step": true,
"logging_steps": 5,
"logging_nan_inf_filter": true,
"save_strategy": "steps",
"save_steps": 50,
"save_total_limit": 2,
"save_safetensors": true,
"save_on_each_node": true,
"save_only_model": false,
"no_cuda": false,
"use_cpu": false,
"use_mps_device": false,
"seed": 42,
"data_seed": null,
"jit_mode_eval": false,
"use_ipex": false,
"bf16": true,
"fp16": false,
"fp16_opt_level": "O1",
"half_precision_backend": "auto",
"bf16_full_eval": false,
"fp16_full_eval": false,
"tf32": null,
"local_rank": 0,
"ddp_backend": "nccl",
"tpu_num_cores": null,
"tpu_metrics_debug": false,
"debug": [],
"dataloader_drop_last": false,
"eval_steps": 50,
"dataloader_num_workers": 1,
"past_index": -1,
"run_name": "/mnt/workspace/output/qwen-1_8b/v0-20240128-024328",
"disable_tqdm": false,
"remove_unused_columns": false,
"label_names": null,
"load_best_model_at_end": true,
"metric_for_best_model": "loss",
"greater_is_better": false,
"ignore_data_skip": false,
"fsdp": [],
"fsdp_min_num_params": 0,
"fsdp_config": {
"min_num_params": 0,
"xla": false,
"xla_fsdp_grad_ckpt": false
},
"fsdp_transformer_layer_cls_to_wrap": null,
"deepspeed": null,
"label_smoothing_factor": 0.0,
"optim": "adamw_torch",
"optim_args": null,
"adafactor": false,
"group_by_length": false,
"length_column_name": "length",
"report_to": [
"tensorboard"
],
"ddp_find_unused_parameters": null,
"ddp_bucket_cap_mb": null,
"ddp_broadcast_buffers": null,
"dataloader_pin_memory": true,
"dataloader_persistent_workers": false,
"skip_memory_metrics": true,
"use_legacy_prediction_loop": false,
"push_to_hub": false,
"resume_from_checkpoint": null,
"hub_model_id": "qwen-1_8b-lora",
"hub_strategy": "every_save",
"hub_token": null,
"hub_private_repo": true,
"hub_always_push": false,
"gradient_checkpointing": true,
"gradient_checkpointing_kwargs": null,
"include_inputs_for_metrics": false,
"fp16_backend": "auto",
"push_to_hub_model_id": null,
"push_to_hub_organization": null,
"push_to_hub_token": null,
"mp_parameters": "",
"auto_find_batch_size": false,
"full_determinism": false,
"torchdynamo": null,
"ray_scope": "last",
"ddp_timeout": 1800,
"torch_compile": false,
"torch_compile_backend": null,
"torch_compile_mode": null,
"dispatch_batches": null,
"split_batches": false,
"include_tokens_per_second": false,
"include_num_input_tokens_seen": false,
"neftune_noise_alpha": null,
"sortish_sampler": true,
"predict_with_generate": false,
"generation_max_length": null,
"generation_num_beams": null,
"generation_config": "GenerationConfig {\n \"chat_format\": \"raw\",\n \"do_sample\": true,\n \"eos_token_id\": 151643,\n \"max_new_tokens\": 2048,\n \"pad_token_id\": 151643,\n \"stop_words_ids\": [\n [\n 151643\n ]\n ],\n \"temperature\": 0.3,\n \"top_k\": 20,\n \"top_p\": 0.7\n}\n",
"train_sampler_random": true,
"push_hub_strategy": "push_best",
"acc_strategy": "token",
"additional_saved_files": [],
"distributed_state": "Distributed environment: NO\nNum processes: 1\nProcess index: 0\nLocal process index: 0\nDevice: cuda\n",
"_n_gpu": 1,
"__cached__setup_devices": "device(type='cuda', index=0)",
"deepspeed_plugin": null
}
示例代码
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import torch
from swift.llm import (DatasetName, InferArguments, ModelType, SftArguments,infer_main, sft_main, app_ui_main, merge_lora_main)
# sft
model_type = ModelType.qwen_1_8b
sft_args = SftArguments(model_type=model_type,sft_type='lora',train_dataset_sample=2000,dataset=[DatasetName.jd_sentiment_zh],output_dir='output')
result = sft_main(sft_args)
best_model_checkpoint = result['best_model_checkpoint']
print(f'best_model_checkpoint: {best_model_checkpoint}')
torch.cuda.empty_cache()
# inference
inf_args = InferArguments(ckpt_dir=best_model_checkpoint,load_dataset_config=True,do_sample=False,max_new_tokens=5)
result = infer_main(inf_args)
torch.cuda.empty_cache()
推理效果
[PROMPT]Task: Sentiment Classification Sentence: 擦玻璃很好、就是太小了 Category: negative, positive Output:
[OUTPUT]positive positive negative
[LABELS]positive
[PROMPT]Task: Sentiment Classification Sentence: 店家太不负责任了，衣服质量太差劲了，和图片上的不一样 Category: negative, positive Output:
[OUTPUT]negative negative negative negative
[LABELS]negative
[PROMPT]Task: Sentiment Classification Sentence: 送国际友人挺好的，不错不错！ Category: negative, positive Output:
[OUTPUT]positive positive positive
[LABELS]positive
[PROMPT]Task: Sentiment Classification Sentence: 很好,装好一定很漂亮 Category: negative, positive Output:
[OUTPUT]positive positive positive
[LABELS]positive
[PROMPT]Task: Sentiment Classification Sentence: 东西给你退回去了，你要黑我钱！！！ Category: negative, positive Output:
[OUTPUT]negative negative negative negative
[LABELS]negative
[PROMPT]Task: Sentiment Classification Sentence: 送货很快，书是正品，买书一直京东是首选！ Category: negative, positive Output:
[OUTPUT]positive positive negative
[LABELS]positive
[PROMPT]Task: Sentiment Classification Sentence: 口感相当的好 都想买第二次了 Category: negative, positive Output:
[OUTPUT]positive positive positive
[LABELS]positive
[PROMPT]Task: Sentiment Classification Sentence: 硅胶味道太重，样子与图片差距太大 Category: negative, positive Output:
[OUTPUT]negative negative negative negative
[LABELS]negative
[PROMPT]Task: Sentiment Classification Sentence: 很伤心，买了放到三星n4尽然不能用，客服各种推 Category: negative, positive Output:
[OUTPUT]negative negative negative negative
[LABELS]negative
[PROMPT]Task: Sentiment Classification Sentence: 质量不错，大小合适，应当是正品！但是我买的是黑灰，发来的却是纯黑，懒得换了，给个差评，希望以后改进！ Category: negative, positive Output:
[OUTPUT]positive negative negative negative
[LABELS]negative
评论