Metadata-Version: 2.4
Name: metrik-ai
Version: 0.3.0
Summary: Metrik — the AI systems engineering platform. Measure, trace, and optimize transformer models.
Project-URL: Homepage, https://github.com/Asmodeus14/Metrik
Project-URL: Repository, https://github.com/Asmodeus14/Metrik
Project-URL: Issues, https://github.com/Asmodeus14/Metrik/issues
Project-URL: Changelog, https://github.com/Asmodeus14/Metrik/blob/main/CHANGELOG.md
Author: The Metrik Authors
License-Expression: Apache-2.0
Keywords: benchmarking,llm,quantization,reproducibility,transformers
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Typing :: Typed
Requires-Python: >=3.11
Requires-Dist: metrik-cli==0.3.0
Provides-Extra: all
Requires-Dist: metrik-profiler[torch]==0.3.0; extra == 'all'
Requires-Dist: metrik-torch[torch]==0.3.0; extra == 'all'
Requires-Dist: transformers>=4.40; extra == 'all'
Provides-Extra: torch
Requires-Dist: metrik-profiler[torch]==0.3.0; extra == 'torch'
Requires-Dist: metrik-torch[torch]==0.3.0; extra == 'torch'
Requires-Dist: transformers>=4.40; extra == 'torch'
Description-Content-Type: text/markdown

# Metrik

**The AI systems engineering platform.** Measure, trace, and optimize transformer models.

```console
$ uv tool install metrik-ai        # or: pip install metrik-ai
$ metrik explore ./models/SmolLM2-135M
```

Every claim Metrik makes traces back to the inputs that produced it. Numbers are labelled
**analytic**, **measured**, or **declared**, and anything that could not be determined is named
with a reason rather than estimated.

## What works today

| Command | What it does |
|---|---|
| `metrik explore <model>` | Architecture, parameters, and memory — no forward pass, no torch |
| `metrik viz <artifact>` | Charts in the terminal, or one self-contained HTML file |
| `metrik store` | Content-addressed artifacts, with full lineage |
| `metrik runs` | What ran, including what failed |
| `metrik bench --check` | What isolation this machine has — and what it will refuse to run |

`metrik explore` reads safetensors headers and `config.json`. It installs and runs on a laptop
with no ML stack at all, in about two seconds for a 270 MB model.

For an observed module tree — real module classes, weight tying by storage identity — install
the extra:

```console
$ uv tool install "metrik-ai[torch]"
$ metrik explore ./models/SmolLM2-135M --loader torch
```

## What does not work yet

Profiling, optimization, and benchmark scoring. See [ROADMAP.md](../../ROADMAP.md). The
benchmark sandbox **refuses** rather than running model-generated code unsandboxed, which is
not revisitable — see [ADR 019](../../docs/planning/adr/019-benchmark-sandbox.md).

Apache-2.0.
