aquin 0.4.0

Stream live training metrics, run model diffs, and inspect datasets — all from your training script.

aquin

Python SDK for Aquin — stream live training metrics, run model diffs, and inspect datasets directly from your training script.

Install

pip install aquin

If you need PyTorch helpers (activation tracking etc.):

pip install "aquin[torch]"

Usage

Training session

import aquin

session = aquin.attach(model, optimizer, api_key="aq-...")

for step in range(num_steps):
    loss = train_step()
    session.step(loss, epoch=epoch, batch=batch, total_batches=len(dataloader))

session.stop()

Open aquin.app/app → Training tab → connect with your API key to watch metrics stream in live.

Model diff

Compare a fine-tuned checkpoint against its base model:

diff = aquin.ModelDiff(
    base="meta-llama/Llama-3.2-1B",
    ft_path="./my-checkpoint.pt",
    api_key="aq-...",
)
result = diff.run(prompts=["Tell me about X", "Explain Y"])

SAE feature diff

Inspect which SAE features changed most between base and fine-tuned:

sae = aquin.SAEDiff(
    base="meta-llama/Llama-3.2-1B",
    ft_path="./my-checkpoint.pt",
    api_key="aq-...",
)
result = sae.run(prompts=["Tell me about X"])

Dataset inspection

Scan a dataset for PII, toxicity, bias, synthetic content, and compliance:

ds = aquin.DataSession(
    dataset_path="./training_data.jsonl",
    api_key="aq-...",
)
results = ds.inspect()

Supports .jsonl, .json, .csv, and .tsv.

API key

Get your key at aquin.app.

Links

• Dashboard
• Docs