Quickstart — your first audit-ready training run

This guide walks you through installing cognify-sdk, instrumenting a training script, and exporting your first compliance package. Estimated time: 5 minutes.

Installation

Install cognify-sdk from PyPI. Python 3.8+ required.

$ pip install cognify-sdk==2.1.4

# Or with extras for your framework:
$ pip install "cognify-sdk[hf,pytorch]"==2.1.4

Initialization

Create a Cognify API key from your workspace settings. Then initialize at the top of your training script:

import os
import cognify

cognify.init(
    workspace="your-workspace-slug",
    api_key=os.environ["COGNIFY_API_KEY"],
    run_name="my-first-ft-run",
    tags=["experiment", "clinical-nlp"]
)

Instrumenting your training loop

Wrap your dataset registration and training call. For Hugging Face Trainer, the SDK auto-hooks into checkpoint events.

from transformers import Trainer, TrainingArguments
from datasets import load_dataset

# Register your dataset — Cognify hashes and versions it
train_ds = load_dataset("my_org/clinical-ner-v2")
cognify.dataset(
    name="clinical-ner-train",
    source=train_ds,
    split="train"
)

training_args = TrainingArguments(
    output_dir="./checkpoints",
    num_train_epochs=3,
    per_device_train_batch_size=16,
    learning_rate=2e-5,
    save_steps=500,
)

trainer = Trainer(model=model, args=training_args, train_dataset=train_ds["train"])
trainer.train()  # Checkpoints auto-logged

Viewing the lineage graph

After training completes, view your run in the Cognify dashboard at app.fyntuneq.com/runs/my-first-ft-run. The lineage graph shows:

• Dataset snapshot with SHA-256 hash and record count
• Training configuration (hyperparameters, hardware)
• Checkpoint sequence with step numbers and loss values
• Eval results (if eval dataset registered)

Exporting your first audit package

Once your run completes, export a compliance package:

import cognify

pkg = cognify.export_audit(
    run_name="my-first-ft-run",
    formats=["pdf", "json"],
    output_dir="./audit_packages"
)
print(f"Exported to: {pkg.output_path}")

The package includes a model card, data provenance report, training configuration attestation, and eval benchmark summary (if eval was configured). Share the PDF with your compliance team — it contains everything needed for model sign-off.