Strikes has early support for integrating with the transformers and derived libraries, allowing you to track and visualize model training directly in the platform. This is accomplished via the DreadnodeCallback, a drop-in callback for the Trainer class.

​
Installation

Make sure you have both dreadnode and transformers installed: 
pip install -U dreadnode transformers datasets

​
Tracking a Transformers Training Run

Below is a minimal example of using the DreadnodeCallback with Hugging Face’s Trainer: 
import dreadnode as dn
from datasets import load_dataset
from transformers import AutoModelForSequenceClassification, AutoTokenizer, Trainer, TrainingArguments
from dreadnode.integrations.transformers import DreadnodeCallback

# Configure Strikes (replace with your API key)
dn.configure(token="<YOUR API KEY>")

# Load and preprocess dataset
dataset = load_dataset("glue", "sst2")
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
dataset["train"] = dataset["train"].select(range(1000))
dataset["validation"] = dataset["validation"].select(range(1000))

def preprocess_function(examples):
    return tokenizer(examples["sentence"], truncation=True, padding="max_length")

tokenized_datasets = dataset.map(preprocess_function, batched=True)

# Load model
model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels=2)

# Define training arguments
training_args = TrainingArguments(
    output_dir="./results",
    learning_rate=2e-5,
    per_device_train_batch_size=6,
    per_device_eval_batch_size=6,
    num_train_epochs=5,
    weight_decay=0.01,
    eval_strategy="steps",
    eval_steps=5,
    load_best_model_at_end=False,
    push_to_hub=False,
    run_name="distilbert-sst2-demo",
)

# Initialize Trainer with DreadnodeCallback
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
    tokenizer=tokenizer,
    callbacks=[DreadnodeCallback(project="training")],
)

# Train and evaluate
trainer.train()
trainer.evaluate()
  • The DreadnodeCallback automatically logs metrics (loss, accuracy, etc.), hyperparameters, and run metadata to Strikes.
  • You can view your training progress and compare runs in the Strikes UI.
  • All data is associated with your project for easy organization.