Dev (#117)
This commit is contained in:
xzw
GitHub
commit
412c80aef3
@ -44,17 +44,16 @@
|
||||
## 数据集去重
|
||||
结合绝对匹配以及模糊匹配(Simhash)算法，对数据集进行去重以提升微调模型的效果。在确保数据集的高质量的同时，通过调整阈值减少因错误匹配而丢失重要数据的风险。
|
||||
|
||||
Simhash算法
|
||||
**Simhash算法介绍**
|
||||
Simhash（相似性哈希）是一种用于检测大量数据中相似或重复项的算法。它通过将文本转换为一组数值指纹来工作，这些指纹对相似的文本具有高度的相似性。Simhash算法对于处理文本数据特别有效，尤其是在处理大量数据时。
|
||||
|
||||
实现步骤：
|
||||
文本预处理：将文本数据转换为适合Simhash处理的格式。这可能包括分词、去除停用词、词干提取等。
|
||||
|
||||
生成Simhash指纹：对预处理后的文本应用Simhash算法，生成一组数值指纹。每个指纹代表文本内容的一个哈希值。
|
||||
|
||||
比较指纹：通过比较哈希值的相似性来识别重复或相似的记录。Simhash的特点是即使在文本有少量差异时，生成的哈希值也具有较高的相似性。
|
||||
|
||||
确定阈值：设置一个相似性阈值，只有当两个指纹的相似度超过这个阈值时，才认为它们代表相似或重复的记录。
|
||||
|
||||
处理相似记录：对于被标记为相似的记录，可以进一步人工审查或自动合并，以消除重复。
|
||||
**Simhash实现步骤**
|
||||
*文本预处理：将文本数据转换为适合Simhash处理的格式。这可能包括分词、去除停用词、词干提取等。
|
||||
*生成Simhash指纹：对预处理后的文本应用Simhash算法，生成一组数值指纹。每个指纹代表文本内容的一个哈希值。
|
||||
*比较指纹：通过比较哈希值的相似性来识别重复或相似的记录。Simhash的特点是即使在文本有少量差异时，生成的哈希值也具有较高的相似性。
|
||||
*确定阈值：设置一个相似性阈值，只有当两个指纹的相似度超过这个阈值时，才认为它们代表相似或重复的记录。
|
||||
*处理相似记录：对于被标记为相似的记录，可以进一步人工审查或自动合并，以消除重复。
|
||||
|
||||
## 用法
|
||||
### deduplicate.py
|
||||
`deduplicate.py` 用于将datasets下以模型命名的文件夹下(例如：'datasets/qwen').json数据进行去重，输出去重后的数据到 `datasets/qwen/dedup` 文件夹下。
|
||||
@ -7552,9 +7552,6 @@
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"conversation": []
|
||||
},
|
||||
{
|
||||
"conversation": [
|
||||
{
|
||||
@ -8540,9 +8537,6 @@
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"conversation": []
|
||||
},
|
||||
{
|
||||
"conversation": [
|
||||
{
|
||||
@ -13389,9 +13383,6 @@
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"conversation": []
|
||||
},
|
||||
{
|
||||
"conversation": [
|
||||
{
|
||||
@ -19973,9 +19964,6 @@
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"conversation": []
|
||||
},
|
||||
{
|
||||
"conversation": [
|
||||
{
|
||||
|
||||
@ -1,12 +1,25 @@
|
||||
import json
|
||||
|
||||
# 打开JSON文件并读取其内容
|
||||
with open('/root/Emollm/datasets/multi_turn_dataset_2.json', 'rt', encoding='utf-8') as file:
|
||||
|
||||
# file_name = 'multi_turn_dataset_1.json'
|
||||
# file_name = 'multi_turn_dataset_2.json'
|
||||
# file_name = 'data_pro.json'
|
||||
file_name = 'data.json'
|
||||
|
||||
with open(f'/root/StableCascade/emollm2/EmoLLM/datasets/{file_name}', 'rt', encoding='utf-8') as file:
|
||||
data = json.load(file)
|
||||
|
||||
n = 0
|
||||
for i in data:
|
||||
i['conversation'][0]['system'] = "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
|
||||
try:
|
||||
i['conversation'][0]['system'] = "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
except:
|
||||
print(n,i) # 4 empty lines in data.json 425 483 742 1120
|
||||
n+=1
|
||||
|
||||
with open('output2.json', 'wt', encoding='utf-8') as file:
|
||||
with open(f'processed_{file_name}', 'wt', encoding='utf-8') as file:
|
||||
json.dump(data, file, ensure_ascii=False, indent=4)
|
||||
|
||||
print(data[0])
|
||||
34
datasets/processed/process_merge.py
Normal file
34
datasets/processed/process_merge.py
Normal file
@ -0,0 +1,34 @@
|
||||
import os
|
||||
import json
|
||||
|
||||
# 设置目录路径，这里假设你的JSON文件都在当前目录下的directory_path文件夹中
|
||||
directory_path = './'
|
||||
|
||||
# 初始化一个空列表，用于存储所有JSON文件的数据
|
||||
combined_list = []
|
||||
|
||||
# 遍历指定目录下的所有文件
|
||||
for filename in os.listdir(directory_path):
|
||||
# 检查文件扩展名是否为.json
|
||||
if filename.endswith('.json'):
|
||||
# 构建文件的完整路径
|
||||
file_path = os.path.join(directory_path, filename)
|
||||
|
||||
# 打开并读取JSON文件
|
||||
with open(file_path, 'r', encoding='utf-8') as json_file:
|
||||
# 加载JSON文件的内容
|
||||
data = json.load(json_file)
|
||||
|
||||
# 将读取到的数据添加到combined_list中
|
||||
# 假设每个JSON文件包含的是一个列表，如果不是，可以根据实际情况调整
|
||||
if isinstance(data, list):
|
||||
combined_list.extend(data)
|
||||
else:
|
||||
combined_list.append(data)
|
||||
|
||||
# 打印合并后的列表 very large and slow
|
||||
# print(combined_list)
|
||||
|
||||
# 如果需要，可以将合并后的列表保存到一个新的JSON文件中
|
||||
with open('combined_data.json', 'w', encoding='utf-8') as combined_json_file:
|
||||
json.dump(combined_list, combined_json_file, ensure_ascii=False, indent=4)
|
||||
@ -0,0 +1,27 @@
|
||||
import json
|
||||
|
||||
# 打开JSON文件并读取其内容
|
||||
# file_name = 'single_turn_dataset_1.json'
|
||||
file_name = 'single_turn_dataset_2.json'
|
||||
with open(f'/root/StableCascade/emollm2/EmoLLM/datasets/{file_name}', 'rt', encoding='utf-8') as file:
|
||||
format1_data = json.load(file)
|
||||
|
||||
system = "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
|
||||
# 转换为格式2的数据
|
||||
format2_data = []
|
||||
for item in format1_data:
|
||||
conversation = {
|
||||
"system": system,
|
||||
"input": item["prompt"],
|
||||
"output": item["completion"]
|
||||
}
|
||||
format2_data.append({"conversation": [conversation]})
|
||||
|
||||
# 将转换后的数据转换为JSON格式
|
||||
|
||||
|
||||
with open(f'./processed_{file_name}', 'wt', encoding='utf-8') as file:
|
||||
json.dump(format2_data, file, ensure_ascii=False, indent=4)
|
||||
|
||||
print(format2_data[0])
|
||||
203
xtuner_config/internlm2_7b_base_qlora_e3.py
Normal file
203
xtuner_config/internlm2_7b_base_qlora_e3.py
Normal file
@ -0,0 +1,203 @@
|
||||
# Copyright (c) OpenMMLab. All rights reserved.
|
||||
import torch
|
||||
from datasets import load_dataset
|
||||
from mmengine.dataset import DefaultSampler
|
||||
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
|
||||
LoggerHook, ParamSchedulerHook)
|
||||
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
|
||||
from peft import LoraConfig
|
||||
from torch.optim import AdamW
|
||||
from transformers import (AutoModelForCausalLM, AutoTokenizer,
|
||||
BitsAndBytesConfig)
|
||||
|
||||
from xtuner.dataset import process_hf_dataset
|
||||
from xtuner.dataset.collate_fns import default_collate_fn
|
||||
from xtuner.dataset.map_fns import template_map_fn_factory
|
||||
from xtuner.engine import DatasetInfoHook, EvaluateChatHook
|
||||
from xtuner.model import SupervisedFinetune
|
||||
from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
|
||||
|
||||
#######################################################################
|
||||
# PART 1 Settings #
|
||||
#######################################################################
|
||||
# Model
|
||||
# pretrained_model_name_or_path = '/root/share/model_repos/internlm2-chat-7b'
|
||||
pretrained_model_name_or_path = '/root/share/model_repos/internlm2-base-7b'
|
||||
|
||||
# Data
|
||||
# data_path = 'merge.json'
|
||||
data_path ='/root/StableCascade/emollm2/EmoLLM/datasets/processed/combined_data.json'
|
||||
|
||||
# https://github.com/InternLM/xtuner/blob/main/xtuner/utils/templates.py#L24C25-L24C25
|
||||
prompt_template = PROMPT_TEMPLATE.internlm2_chat # there is No internlm2_base
|
||||
|
||||
max_length = 2048
|
||||
pack_to_max_length = True
|
||||
|
||||
# Scheduler & Optimizer
|
||||
|
||||
# batch_size = 8 # per_device
|
||||
# accumulative_counts = 2
|
||||
batch_size = 16 # per_device
|
||||
accumulative_counts = 1
|
||||
|
||||
dataloader_num_workers = 0
|
||||
max_epochs = 3
|
||||
optim_type = AdamW
|
||||
lr = 2e-4
|
||||
betas = (0.9, 0.999)
|
||||
weight_decay = 0
|
||||
max_norm = 1 # grad clip
|
||||
warmup_ratio = 0.03
|
||||
|
||||
# Evaluate the generation performance during the training
|
||||
evaluation_freq = 500
|
||||
# SYSTEM = "现在你是一个心理专家，我有一些心理问题，请你用专业的知识帮我解决。"
|
||||
SYSTEM = "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
evaluation_inputs = [
|
||||
'我最近总是感到很焦虑，尤其是在学业上。我有个特别崇拜的同学，他好像在各方面都比我优秀，我总觉得自己怎么努力也追不上他，这让我压力特别大。', '我知道应该理性看待，但就是忍不住会去比较。我甚至晚上会因为这个睡不着觉，总想着怎样才能像他那样出色。'
|
||||
]
|
||||
|
||||
#######################################################################
|
||||
# PART 2 Model & Tokenizer #
|
||||
#######################################################################
|
||||
tokenizer = dict(
|
||||
type=AutoTokenizer.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
padding_side='right')
|
||||
|
||||
model = dict(
|
||||
type=SupervisedFinetune,
|
||||
llm=dict(
|
||||
type=AutoModelForCausalLM.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
torch_dtype=torch.float16,
|
||||
quantization_config=dict(
|
||||
type=BitsAndBytesConfig,
|
||||
load_in_4bit=True,
|
||||
load_in_8bit=False,
|
||||
llm_int8_threshold=6.0,
|
||||
llm_int8_has_fp16_weight=False,
|
||||
bnb_4bit_compute_dtype=torch.float16,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
bnb_4bit_quant_type='nf4')),
|
||||
lora=dict(
|
||||
type=LoraConfig,
|
||||
r=64,
|
||||
lora_alpha=16,
|
||||
lora_dropout=0.1,
|
||||
bias='none',
|
||||
task_type='CAUSAL_LM'))
|
||||
|
||||
#######################################################################
|
||||
# PART 3 Dataset & Dataloader #
|
||||
#######################################################################
|
||||
alpaca_en = dict(
|
||||
type=process_hf_dataset,
|
||||
dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)),
|
||||
tokenizer=tokenizer,
|
||||
max_length=max_length,
|
||||
dataset_map_fn=None,
|
||||
template_map_fn=dict(
|
||||
type=template_map_fn_factory, template=prompt_template),
|
||||
remove_unused_columns=True,
|
||||
shuffle_before_pack=True,
|
||||
pack_to_max_length=pack_to_max_length)
|
||||
|
||||
train_dataloader = dict(
|
||||
batch_size=batch_size,
|
||||
num_workers=dataloader_num_workers,
|
||||
dataset=alpaca_en,
|
||||
sampler=dict(type=DefaultSampler, shuffle=True),
|
||||
collate_fn=dict(type=default_collate_fn))
|
||||
|
||||
#######################################################################
|
||||
# PART 4 Scheduler & Optimizer #
|
||||
#######################################################################
|
||||
# optimizer
|
||||
optim_wrapper = dict(
|
||||
type=AmpOptimWrapper,
|
||||
optimizer=dict(
|
||||
type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
|
||||
clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
|
||||
accumulative_counts=accumulative_counts,
|
||||
loss_scale='dynamic',
|
||||
dtype='float16')
|
||||
|
||||
# learning policy
|
||||
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501
|
||||
param_scheduler = [
|
||||
dict(
|
||||
type=LinearLR,
|
||||
start_factor=1e-5,
|
||||
by_epoch=True,
|
||||
begin=0,
|
||||
end=warmup_ratio * max_epochs,
|
||||
convert_to_iter_based=True),
|
||||
dict(
|
||||
type=CosineAnnealingLR,
|
||||
eta_min=0.0,
|
||||
by_epoch=True,
|
||||
begin=warmup_ratio * max_epochs,
|
||||
T_max=max_epochs,
|
||||
convert_to_iter_based=True)
|
||||
]
|
||||
|
||||
# train, val, test setting
|
||||
train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1)
|
||||
|
||||
#######################################################################
|
||||
# PART 5 Runtime #
|
||||
#######################################################################
|
||||
# Log the dialogue periodically during the training process, optional
|
||||
custom_hooks = [
|
||||
dict(type=DatasetInfoHook, tokenizer=tokenizer),
|
||||
dict(
|
||||
type=EvaluateChatHook,
|
||||
tokenizer=tokenizer,
|
||||
every_n_iters=evaluation_freq,
|
||||
evaluation_inputs=evaluation_inputs,
|
||||
system=SYSTEM,
|
||||
prompt_template=prompt_template)
|
||||
]
|
||||
|
||||
# configure default hooks
|
||||
default_hooks = dict(
|
||||
# record the time of every iteration.
|
||||
timer=dict(type=IterTimerHook),
|
||||
# print log every 100 iterations.
|
||||
logger=dict(type=LoggerHook, interval=10),
|
||||
# enable the parameter scheduler.
|
||||
param_scheduler=dict(type=ParamSchedulerHook),
|
||||
# save checkpoint per epoch.
|
||||
checkpoint=dict(type=CheckpointHook, interval=1),
|
||||
# set sampler seed in distributed evrionment.
|
||||
sampler_seed=dict(type=DistSamplerSeedHook),
|
||||
)
|
||||
|
||||
# configure environment
|
||||
env_cfg = dict(
|
||||
# whether to enable cudnn benchmark
|
||||
cudnn_benchmark=False,
|
||||
# set multi process parameters
|
||||
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
|
||||
# set distributed parameters
|
||||
dist_cfg=dict(backend='nccl'),
|
||||
)
|
||||
|
||||
# set visualizer
|
||||
visualizer = None
|
||||
|
||||
# set log level
|
||||
log_level = 'INFO'
|
||||
|
||||
# load from which checkpoint
|
||||
load_from = None
|
||||
|
||||
# whether to resume training from the loaded checkpoint
|
||||
resume = False
|
||||
|
||||
# Defaults to use random seed and disable `deterministic`
|
||||
randomness = dict(seed=None, deterministic=False)
|
||||
@ -1,11 +1,23 @@
|
||||
datasets==2.16.1
|
||||
deepspeed==0.13.1
|
||||
einops==0.7.0
|
||||
flash_attn==2.5.0
|
||||
mmengine==0.10.2
|
||||
openxlab==0.0.34
|
||||
peft==0.7.1
|
||||
sentencepiece==0.1.99
|
||||
torch==2.1.2
|
||||
transformers==4.36.2
|
||||
xtuner==0.1.11
|
||||
|
||||
# modified version
|
||||
# xtuner==0.1.11
|
||||
# mmengine==0.10.2
|
||||
mmengine==0.10.3
|
||||
xtuner==0.1.15
|
||||
|
||||
# flash_attn==2.5.0 # build is very slow about 2 hours?
|
||||
|
||||
# method 1: https://github.com/Dao-AILab/flash-attention/releases
|
||||
# flash_attn-2.5.0+cu122torch2.1cxx11abiTRUE-cp310-cp310-linux_x86_64.whl
|
||||
# method 2:
|
||||
# pip install /root/share/wheels/flash_attn-2.4.2+cu118torch2.0cxx11abiTRUE-cp310-cp310-linux_x86_64.whl
|
||||
|
||||
# mpi4py==3.1.5 # conda install mpi4py
|
||||
|
||||
Loading…
Reference in New Issue
Block a user