Skip to main content

Fine tune LLM on database

Configure your production system

note

If you would like to use the production features of SuperDuperDB, then you should set the relevant connections and configurations in a configuration file. Otherwise you are welcome to use "development" mode to get going with SuperDuperDB quickly.

import os

os.makedirs('.superduperdb', exist_ok=True)
os.environ['SUPERDUPERDB_CONFIG'] = '.superduperdb/config.yaml'
CFG = '''
data_backend: mongodb://127.0.0.1:27017/documents
artifact_store: filesystem://./artifact_store
cluster:
cdc:
strategy: null
uri: ray://127.0.0.1:20000
compute:
uri: ray://127.0.0.1:10001
vector_search:
backfill_batch_size: 100
type: in_memory
uri: http://127.0.0.1:21000
'''
with open(os.environ['SUPERDUPERDB_CONFIG'], 'w') as f:
f.write(CFG)

Start your cluster

note

Starting a SuperDuperDB cluster is useful in production and model development if you want to enable scalable compute, access to the models by multiple users for collaboration, monitoring.

If you don't need this, then it is simpler to start in development mode.

!python -m superduperdb local-cluster up        

Connect to SuperDuperDB

note

Note that this is only relevant if you are running SuperDuperDB in development mode. Otherwise refer to "Configuring your production system".

from superduperdb import superduper

db = superduper('mongodb://localhost:27017/documents')
!pip install transformers torch accelerate trl peft datasets

Get LLM Finetuning Data

The following are examples of training data in different formats.

from datasets import load_dataset
from superduperdb.base.document import Document
dataset_name = "timdettmers/openassistant-guanaco"
dataset = load_dataset(dataset_name)

train_dataset = dataset["train"]
eval_dataset = dataset["test"]

train_documents = [
Document({**example, "_fold": "train"})
for example in train_dataset
]
eval_documents = [
Document({**example, "_fold": "valid"})
for example in eval_dataset
]

datas = train_documents + eval_documents

We can define different training parameters to handle this type of data.

# Function for transformation after extracting data from the database
transform = None
key = ('text')
training_kwargs=dict(dataset_text_field="text")

Example input_text and output_text

data = datas[0]
input_text, output_text = data["text"].rsplit("### Assistant: ", maxsplit=1)
input_text += "### Assistant: "
output_text = output_text.rsplit("### Human:")[0]
print("Input: --------------")
print(input_text)
print("Response: --------------")
print(output_text)

Setup simple tables or collections

# If our data is in a format natively supported by MongoDB, we don't need to do anything.
from superduperdb.backends.mongodb import Collection

table_or_collection = Collection('documents')
select = table_or_collection.find({})

Insert simple data

In order to create data, we need to create a Schema for encoding our special Datatype column(s) in the databackend.

from superduperdb import Document

ids, _ = db.execute(table_or_collection.insert_many(datas))

Select a Model

model_name = "facebook/opt-125m"
model_kwargs = dict()
tokenizer_kwargs = dict()

# or
# model_name = "mistralai/Mistral-7B-Instruct-v0.2"
# token = "hf_xxxx"
# model_kwargs = dict(token=token)
# tokenizer_kwargs = dict(token=token)

Build A Trainable LLM

Create an LLM Trainer for training

The parameters of this LLM Trainer are basically the same as transformers.TrainingArguments, but some additional parameters have been added for easier training setup.

from superduperdb.ext.transformers import LLM, LLMTrainer
trainer = LLMTrainer(
identifier="llm-finetune-trainer",
output_dir="output/finetune",
overwrite_output_dir=True,
num_train_epochs=3,
save_total_limit=3,
logging_steps=10,
evaluation_strategy="steps",
save_steps=100,
eval_steps=100,
per_device_train_batch_size=1,
per_device_eval_batch_size=1,
gradient_accumulation_steps=2,
max_seq_length=512,
key=key,
select=select,
transform=transform,
training_kwargs=training_kwargs,
)
trainer.use_lora = True        

Create a trainable LLM model and add it to the database, then the training task will run automatically.

llm = LLM(
identifier="llm",
model_name_or_path=model_name,
trainer=trainer,
model_kwargs=model_kwargs,
tokenizer_kwargs=tokenizer_kwargs,
)

db.apply(llm)

Load the trained model

There are two methods to load a trained model:

  • Load the model directly: This will load the model with the best metrics (if the transformers' best model save strategy is set) or the last version of the model.
  • Use a specified checkpoint: This method downloads the specified checkpoint, then initializes the base model, and finally merges the checkpoint with the base model. This approach supports custom operations such as resetting flash_attentions, model quantization, etc., during initialization.
llm = db.load("model", "llm")        
llm.predict_one(input_text, max_new_tokens=200)