Overview

Aquin Interp Tooling is the open-source CLI and Python SDK from Aquin Labs Private Limited.

Fine-tunes fail in ways loss curves do not explain. AQIT is one place to declare a Recipe, train, EvalGate on your probes, inspect only if the gate fails, and patch cheaply (steer) or retrain from a new data revision.

Same objects in CLI and SDK. No second product for inspect.

Start here

I want to…	Go
Understand the loop	Recipe loop
See how the tree is layered	Architecture
Run the CLI	CLI
Call it from Python	Python SDK
Trace / SAE / steer	Inspect, SAE, patch
GPU kernels and evals	Compute
Local daemon	Engine daemon
Build these docs	Building these docs

30-second loop

pip install -e .
aqit train recipe.yaml
aqit eval custom --prompts probes.jsonl
aqit trace --prompt "Hello" --layer 8

import aqit
 
aqit.train.run("recipe.yaml")
aqit.eval.custom(prompts="probes.jsonl")
aqit.inspect.trace(prompt="Hello", layer=8)
aqit.patch.steer(feature_idx=42, strength=2.0)

Why it exists

1. Declare a Recipe (data, method, gate).
2. Train (tabular linear/boosting or LLM LoRA).
3. EvalGate on your probes: pass or fail.
4. Inspect the run if it failed (trace, SAE, diffs).
5. Patch cheaply (steer) or retrain from a new data revision.

flowchart LR
  R[Recipe] --> T[Train]
  T --> C[Checkpoint]
  C --> G[EvalGate]
  G -->|pass| D[Done]
  G -->|fail| I[Inspect]
  I --> P[Patch / steer]
  P --> R

Public names

Surface	Import / bin	Lives in
Public package	aqit	repo root (cli.py, __init__.py)
Engine package	aquin	interp-engine/ (mapped in setup.py)
CLI	aqit	aqit.cli:main
Alias CLI	aquin	aquin.cli:main

The public import is aqit. The engine is still imported internally as aquin until that rename lands. Aquin the foundation-model framework is a separate product. This tree is what ships today.

License

Apache License 2.0. Copyright 2026 Aquin Labs Private Limited. See LICENSE.