The purpose of multi-turn dialogue command fine-tuning is to enhance the model's ability for multi-turn dialogues.
XTuner supports the use of HuggingFace Hub datasets or custom datasets for SFT (Supervised FineTune). The main difference between them is that when using the HuggingFace Hub dataset, the original data needs to be mapped to the multi-turn dialogue data format defined by XTuner. For custom datasets, it is recommended that users construct the dataset according to the multi-turn dialogue data format.
Since the formats of different datasets vary, the original data needs to be transformed into the multi-turn dialogue data format defined by XTuner. XTuner supports the use of a map function to achieve format mapping. The following example uses the oasst1 dataset to illustrate how to implement data mapping.
The oasst1 dataset format is as follows:
>>> from datasets import load_dataset
>>> ds = load_dataset(path='timdettmers/openassistant-guanaco')
>>> ds['train']
Dataset({
features: ['text'],
num_rows: 9846
})
>>> ds['train'][0]['text']
'### Human: xxx ### Assistant: xxx ###Human: xxx ###Assistant: xxx'It's clear that the oasst1 dataset can not only be used as an incremental pre-training dataset for the model to learn some basic language knowledge, but also, after some processing, serve as a multi-turn dialogue dataset to cultivate the model's multi-turn conversation capabilities. The multi-turn dialogue data format introduces that in the fine-tuning process of multi-turn dialogue instructions, the data format should be:
[{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
},
{
"input": "xxx",
"output": "xxx"
}
]
},
{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
},
{
"input": "xxx",
"output": "xxx"
}
]
}]Therefore, the original data can be mapped to a standard format using the following map function:
# Suppose the function is stored in ./map_fn.py
SYSTEM_OASST1 = '' # oasst1 does not set the system text
def custom_map_fn(example):
r"""
Example before preprocessing:
example['text'] = '### Human: Can you explain xxx'
'### Assistant: Sure! xxx'
'### Human: I didn't understand how xxx'
'### Assistant: It has to do with a process xxx.'
Example after preprocessing:
example['conversation'] = [
{
'input': 'Can you explain xxx',
'output': 'Sure! xxx'
},
{
'input': 'I didn't understand how xxx',
'output': 'It has to do with a process xxx.'
}
]
"""
data = []
for sentence in example['text'].strip().split('###'):
sentence = sentence.strip()
if sentence[:6] == 'Human:':
data.append(sentence[6:].strip())
elif sentence[:10] == 'Assistant:':
data.append(sentence[10:].strip())
if len(data) % 2:
# The last round of conversation solely consists of input
# without any output.
# Discard the input part of the last round, as this part is ignored in
# the loss calculation.
data.pop()
conversation = []
for i in range(0, len(data), 2):
system = SYSTEM_OASST1 if i == 0 else ''
single_turn_conversation = {
'system': system,
'input': data[i],
'output': data[i + 1]}
conversation.append(single_turn_conversation)
return {'conversation': conversation}XTuner provides several ready-to-use configuration files. Users can view them using the following command:
xtuner list-cfg -p internlm-p is used for fuzzy search. If you want to train other models, you can replace internlm with other model names supported by XTuner.
If the provided configuration file does not meet your needs, please export the offered configuration file and make appropriate changes:
xtuner copy-cfg ${CONFIG_NAME} ${SAVE_DIR}For example, use the following command to export the config named internlm_7b_qlora_oasst1_e3 to the current directory:
xtuner copy-cfg internlm_7b_qlora_oasst1_e3 .The config file copied in Step 3 needs to be modified as follows:
- Import the map function
custom_map_fnimplemented in Step 1. - Replace
dataset_map_fnintrain_datasetwithcustom_map_fn. - Adjust the path of the original dataset. You can refer to the user documentation for operations related to
load_dataset.
from xtuner.dataset import process_hf_dataset
from datasets import load_dataset
- from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory
+ from xtuner.dataset.map_fns import template_map_fn_factory
+ from mmengine.config import read_base
+ with read_base():
+ from .map_fn import custom_map_fn
...
#######################################################################
# PART 1 Settings #
#######################################################################
- data_path = 'timdettmers/openassistant-guanaco'
+ data_path = 'path/to/your/data'
...
#######################################################################
# STEP 3 Dataset & Dataloader #
#######################################################################
train_dataset = dict(
type=process_hf_dataset,
dataset=dict(type=load_dataset, path=data_path),
tokenizer=tokenizer,
max_length=max_length,
- dataset_map_fn=oasst1_map_fn,
+ dataset_map_fn=custom_map_fn,
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)
...After modifying the config file, you can execute the 'xtuner/tools/check_custom_dataset.py' script to verify the correct construction of the dataset.
xtuner check-custom-dataset $CONFIG$CONFIG represents the file path of the modified configuration file in Step 4.
When using a custom multi-turn dialogue dataset for command fine-tuning, we recommend constructing the dataset in the multi-turn dialogue data format as defined by XTuner. If the custom dataset format is oasst1 or other formats, you can refer to the section on Using Datasets in HuggingFace Hub.
Prepare your custom data according to the multi-turn dialogue data format defined by XTuner:
[{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
},
{
"input": "xxx",
"output": "xxx"
}
]
},
{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
},
{
"input": "xxx",
"output": "xxx"
}
]
}]xtuner list-cfg -p internlm-p is for fuzzy search. If you want to train other models, you can replace internlm with other model names supported by XTuner.
xtuner copy-cfg internlm_7b_qlora_oasst1_e3 .The config file copied in Step 3 needs to be modified as follows:
- Adjust the path of the original dataset
- Since the dataset format is already in the standard format, set
dataset_map_fnintrain_datasettoNone
from xtuner.dataset import process_hf_dataset
from datasets import load_dataset
- from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory
+ from xtuner.dataset.map_fns import template_map_fn_factory
...
#######################################################################
# PART 1 Settings #
#######################################################################
- data_path = 'timdettmers/openassistant-guanaco'
+ data_path = 'path/to/your/json/data'
...
#######################################################################
# STEP 3 Dataset & Dataloader #
#######################################################################
train_dataset = dict(
type=process_hf_dataset,
- dataset=dict(type=load_dataset, path=data_path),
+ dataset=dict(
+ type=load_dataset, path='json', data_files=dict(train=data_path)),
tokenizer=tokenizer,
max_length=max_length,
- dataset_map_fn=oasst1_map_fn,
+ 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)
...After modifying the config file, you can execute the 'xtuner/tools/check_custom_dataset.py' script to verify the correct construction of the dataset.
xtuner check-custom-dataset $CONFIG$CONFIG represents the file path of the modified configuration file in Step 4.