train_script.py

from datasets import load_dataset
from sentence_transformers import (
    SparseEncoder,
    SparseEncoderTrainer,
    SparseEncoderTrainingArguments,
    SparseEncoderModelCardData,
)
from sentence_transformers.sparse_encoder.losses import SpladeLoss, SparseMultipleNegativesRankingLoss
from sentence_transformers.training_args import BatchSamplers
from sentence_transformers.sparse_encoder.evaluation import SparseNanoBEIREvaluator
from sentence_transformers.sparse_encoder.models import SpladePooling, MLMTransformer, SparseStaticEmbedding
from sentence_transformers.models import Router

import logging

logging.basicConfig(format="%(asctime)s - %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO)

# 1. Load a model to finetune with 2. (Optional) model card data
mlm_transformer = MLMTransformer("prajjwal1/bert-tiny", tokenizer_args={"model_max_length": 512})
splade_pooling = SpladePooling(
    pooling_strategy="max", word_embedding_dimension=mlm_transformer.get_sentence_embedding_dimension()
)

router = Router(
    {
        "query": [SparseStaticEmbedding(tokenizer=mlm_transformer.tokenizer, frozen=False)],
        "document": [mlm_transformer, splade_pooling],
    }
)

model = SparseEncoder(
    modules=[router],
    model_card_data=SparseEncoderModelCardData(
        language="en",
        license="apache-2.0",
        model_name="Inference-free SPLADE BERT-tiny trained on Natural-Questions tuples",
    ),
)

# 3. Load a dataset to finetune on
full_dataset = load_dataset("sentence-transformers/natural-questions", split="train").select(range(100_000))
# full_dataset = full_dataset.map(lambda sample: {"query": {"query": sample["query"]}, "corpus": {"corpus": sample["answer"]}}, remove_columns=["query", "answer"])
dataset_dict = full_dataset.train_test_split(test_size=1_000, seed=12)
train_dataset = dataset_dict["train"]
eval_dataset = dataset_dict["test"]
print(train_dataset)
print(train_dataset[0])

# 4. Define a loss function
loss = SpladeLoss(
    model=model,
    loss=SparseMultipleNegativesRankingLoss(model=model),
    lambda_query=0,
    lambda_corpus=3e-2,
)

# 5. (Optional) Specify training arguments
run_name = "inference-free-splade-bert-tiny-nq-fresh-3e-2-lambda-corpus-1e-3-idf-lr-2e-5-lr"
args = SparseEncoderTrainingArguments(
    # Required parameter:
    output_dir=f"models/{run_name}",
    # Optional training parameters:
    num_train_epochs=1,
    per_device_train_batch_size=64,
    per_device_eval_batch_size=64,
    learning_rate=2e-5,
    learning_rate_mapping={"IDF\.weight": 1e-3},  # Set a higher learning rate for the IDF module
    warmup_ratio=0.1,
    fp16=True,  # Set to False if you get an error that your GPU can't run on FP16
    bf16=False,  # Set to True if you have a GPU that supports BF16
    batch_sampler=BatchSamplers.NO_DUPLICATES,  # MultipleNegativesRankingLoss benefits from no duplicate samples in a batch
    router_mapping=["query", "document"],
    # Optional tracking/debugging parameters:
    eval_strategy="steps",
    eval_steps=200,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=2,
    logging_steps=20,
    run_name=run_name,  # Will be used in W&B if `wandb` is installed
)

# 6. (Optional) Create an evaluator & evaluate the base model
dev_evaluator = SparseNanoBEIREvaluator(dataset_names=["msmarco", "nfcorpus", "nq"], batch_size=128)

# 7. Create a trainer & train
trainer = SparseEncoderTrainer(
    model=model,
    args=args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    loss=loss,
    evaluator=dev_evaluator,
)
trainer.train()

# 8. Evaluate the model performance again after training
dev_evaluator(model)

# 9. Save the trained model
model.save_pretrained(f"models/{run_name}/final")

# 10. (Optional) Push it to the Hugging Face Hub
model.push_to_hub(run_name)