Important
NeMo 2.0 is an experimental feature and currently released in the dev container only: nvcr.io/nvidia/nemo:dev. Please refer to the Migration Guide for information on getting started.
BERT Embedding Models
Sentence-BERT (SBERT) is a modification of the BERT model that is specifically trained to generate semantically meaningful sentence embeddings. The model architecture and pre-training process are detailed in the Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks paper. Similar to BERT, Sentence-BERT utilizes a BERT-based architecture, but it is trained using a Siamese and triplet network structure to derive fixed-sized sentence embeddings that capture semantic information. Sentence-BERT is commonly used to generate high-quality sentence embeddings for various downstream natural language processing tasks, such as semantic textual similarity, clustering, and information retrieval
Data Input for the Sentence-BERT model
The fine-tuning data for the Sentence-BERT (SBERT) model should consist of data instances, each comprising a query, a positive document, and a list of negative documents. Negative mining is not supported in NeMo yet; therefore, data preprocessing should be performed offline before training. The dataset should be in JSON format. For instance, the dataset should have the following structure:
[
{
"query": "Query",
"pos_doc": "Positive",
"neg_doc": ["Negative_1", "Negative_2", ..., "Negative_n"]
},
{
// Next data instance
},
...,
{
// Subsequent data instance
}
]
This format ensures that the fine-tuning data is appropriately structured for training the Sentence-BERT model.
Fine-tuning the Sentence-BERT model
For fine-tuning Sentence-BERT model, you need to initialize the Sentence-BERT model with BERT model
checkpoint. To do so, you should either have a .nemo checkpoint or need to convert a HuggingFace
BERT checkpoint to NeMo (mcore) using the following:
python NeMo/scripts/nlp_language_modeling/convert_bert_hf_to_nemo.py \
--input_name_or_path "intfloat/e5-large-unsupervised" \
--output_path /path/to/output/nemo/file.nemo \
--mcore True \
--precision 32
Then you can fine-tune the sentence-BERT model using the following script:
#!/bin/bash
PROJECT= # wandb project name
NAME= # wandb run name
export WANDB_API_KEY= # your_wandb_key
NUM_DEVICES=1 # number of gpus to train on
CONFIG_PATH="/NeMo/examples/nlp/information_retrieval/conf/"
CONFIG_NAME="megatron_bert_embedding_config"
PATH_TO_NEMO_MODEL= # Path to conveted nemo model from hf
TRAIN_DATASET_PATH= # Path to json dataset
VALIDATION_DATASET_PATH= # Path to validation dataset
SAVE_DIR= # where the checkpoint and logs are saved
mkdir -p $SAVE_DIR
export NVTE_FLASH_ATTN=0
export NVTE_ALLOW_NONDETERMINISTIC_ALGO=0
export NVTE_FUSED_ATTN=0
python NeMo/examples/nlp/information_retrieval/megatron_bert_embedding_finetuning.py \
--config-path=${CONFIG_PATH} \
--config-name=${CONFIG_NAME} \
restore_from_path=${PATH_TO_NEMO_MODEL} \
trainer.devices=${NUM_DEVICES} \
trainer.max_steps=10000 \
trainer.val_check_interval=100 \
trainer.max_epochs=1 \
+trainer.num_sanity_val_steps=0 \
model.mcore_bert=True \
model.post_process=False \
model.global_batch_size=8 \ # should be NUM_DEVICES * model.micro_batch_size
model.micro_batch_size=8 \
model.optim.lr=0.000005 \
model.optim.sched.min_lr=0.00000001 \
model.optim.sched.warmup_steps=100 \
model.encoder_seq_length=512 \
model.tokenizer.library="huggingface" \
model.tokenizer.type="intfloat/e5-large-unsupervised" \
model.data.data_train=${TRAIN_DATASET_PATH} \
model.data.data_validation=${VALIDATION_DATASET_PATH} \
model.data.hard_negatives_to_train=4 \
exp_manager.explicit_log_dir=${SAVE_DIR} \
exp_manager.create_wandb_logger=True \
exp_manager.resume_if_exists=True \
exp_manager.wandb_logger_kwargs.name=${NAME} \
exp_manager.wandb_logger_kwargs.project=${PROJECT}
GPT Embedding Models
Recent work has also shown that it is possible to use Decoder-Only (GPT Style) models to train embedding models. Improving Text Embeddings with Large Language Models is one such recent papers which served as inspiration to implement Decoder-only embedding training in Nemo.
Training a GPT Embedding Model
To train GPT Embedding models we follow a format very similar to the SBERT Embedding training. However, there are a couple of differences. GPT Embedding model training expects a jsonl file in which each line is a json object. Here is a truncated example of data jsonl file:
{"query": "What did ... 1952-2002 period?", "pos_doc": "Morning (2008) ... has changed little.", "neg_doc": "Even though ... sapiens.", "query_id": "q103151", "doc_id": "d14755"}
{"query": "What type of ... passions?", "pos_doc": "Burke was a leading ... upper classes.", "neg_doc": "Writing to a friend ... Government.", "query_id": "q77959", "doc_id": "d11263"}
{"query": "Since 1999, ... progressed at?", "pos_doc": "Commercial solar water ... as of 2007.", "neg_doc": "The potential solar ... acquire.", "query_id": "q16545", "doc_id": "d1883"}
As visible the json object should contain the following fields query, pos_doc, neg_doc, query_id and doc_id. The query_id and doc_id can be any alphanumeric string that uniquely maps to the query string and pos_doc string.
During training, the GPT Embedding model employs LoRA (by default) to learn embeddings for the queries and documents, such that similarity of the query-to-pos_doc are maximized while simultaneously minimizing query-to-neg_doc similarity. LoRA allows us to fine-tune large LLMs such as Mistral 7B model with a relatively small number of training parameters.
An example command to launch a training job is
python3 /NeMo/examples/nlp/information_retrieval/megatron_gpt_embedding_finetuning.py \
exp_manager.exp_dir="PATH_TO_SAVE_LORA_WEIGHTS" \
model.global_batch_size=4 \ # exact choice for global batch size is data dependent typical values are in the range of 32 to 128.
model.micro_batch_size=4 \ # exact choice for micro batch size is GPU memory dependent 2 to 8 are reasonable values.
trainer.devices=1 \ # indicates how many GPUs to use during training per node.
trainer.num_nodes=1 \ # indicates how many nodes to use if multi-node cluster is available
trainer.max_steps=20 \ # how many training steps to run.
model.restore_from_path="PATH_TO_BASE_NEMO_MODEL" \
model.peft.lora_tuning.adapter_dim=16 \ # the low-rank size for lora weights.
model.data.train_ds.file_names=["train.jsonl"]
The full list of possible run arguments is configurable in /examples/nlp/information_retrieval/conf/megatron_gpt_embedder_tuning_config.yaml. By default a trained model file should be generated in here PATH_TO_SAVE_LORA_WEIGHTS/megatron_gpt_peft_lora_tuning/checkpoints/ typically with the extension .nemo.
Inference using a GPT Embedding Model
Once trained, the GPT Embedding Model can be used to generate embeddings for queries and corpus documents. We can launch inference using the following command:
python3 /NeMo/examples/nlp/information_retrieval/megatron_gpt_embedding_generate.py \
model.global_batch_size=4 \
model.micro_batch_size=4 \
trainer.devices=1 \
trainer.num_nodes=1 \
model.restore_from_path="PATH_TO_BASE_NEMO_MODEL" \ # Same base model used at training time.
model.peft.restore_from_path="PATH_TO_SAVE_LORA_WEIGHTS/megatron_gpt_peft_lora_tuning/checkpoints//megatron_gpt_peft_lora_tuning.nemo" \
model.data.test_ds.query_file_names=["test_query.jsonl"] \
model.data.test_ds.doc_file_names=\["test_docs.jsonl"] \
model.data.test_ds.write_embeddings_to_file=True \
model.data.test_ds.output_file_path_prefix="PATH_TO_SAVE_EMEBDDINGS"
The contents of test_queries.jsonl is expected to be in the following format:
{"query": "What do ... quantities?","query_id": "q11600", "doc_id": "d1172"}
{"query": "What are ... subsectors?", "query_id": "q5831", "doc_id": "d577"}
{"query": "Which article ... Government?", "query_id": "q3037", "doc_id": "d336"}
Here, the doc_id field is expected to be the id of the document/passage which is the correct passage for the query. Note that since we are in inference mode, we don’t require query-doc pairs.
The contents of test_docs.jsonl is expected to be in the following format:
{"pos_doc": "Hormones ... vitamin D.", "doc_id": "d823"}
{"pos_doc": "Historically, Victoria ... October 2016.", "doc_id": "d159"}
{"pos_doc": "Exceptional examples ... Warsaw.", "doc_id": "d1084"}
Once again, we show 3 examples form each file. Typically the test_docs.jsonl will contain more items than queries in the test_queries.jsonl.
The inference command will result in two folders
PATH_TO_SAVE_EMBEDDINGS/consumed_samplesX/test_queriesPATH_TO_SAVE_EMBEDDINGS/consumed_samplesX/test_docs
The X in the folder consumed_samplesX is a number denoted number of batches consumed, this is not crucial at test time, but it is useful in training which we will see in the next section. First, let’s take a look at the test_queries.
$> ls PATH_TO_SAVE_EMBEDDINGS/consumed_samplesX/test_queries
query.ids query.npy
$>head -n3 PATH_TO_SAVE_EMBEDDINGS/consumed_samplesX/test_queries/query.ids
q11600
q5831
q3037
query.npy is a numpy pickled array containing rows of query embeddings and the query.ids text file list the id of each embedding in the same order.
Similarly let’s look into the test_docs folder
$> ls PATH_TO_SAVE_EMBEDDINGS/consumed_samplesX/test_doc/
doc.ids doc.npy
$> head -n3 PATH_TO_SAVE_EMBEDDINGS/consumed_samplesX/test_doc/doc.ids
d823
d159
d1084
We can see that test_doc has a similar structure to test_queries but with ids and embeddings of the documents from the test_docs.josnl file. With this setup it is possible to evaluate the performance using metrics like MRR or NDCG.