Steerling-8B: Official Model, Attribution and Evaluation Guide

Editorial update: November 2, 2026. This revision checks the upstream Steerling project and model card. It replaces the earlier unsourced tutorial with a guide to the published implementation and its evaluation boundaries.

Identify the actual project and model

Steerling is a Guide Labs project. The official code repository and Steerling-8B model card identify the base model as guidelabs/steerling-8b. The model card describes an approximately eight-billion-parameter causal diffusion language model with concept decomposition and a 4,096-token context. Its generation process iteratively unmasks tokens rather than using a conventional left-to-right generation loop.

The model card gives an approximately 18 GB VRAM requirement for its documented bfloat16 setup. That is an upstream requirement, not a measurement of memory use or throughput made for this article. Check the exact software and hardware configuration before downloading weights or planning a deployment.

What the implementation lets you inspect

At the reviewed code revision, the project provides concept-contribution analysis, attribution to input tokens and concept interventions. It distinguishes known concepts, discovered features and a reconstruction correction. The repository also states that training-data attribution is not supported in that release.

Treat these outputs as specific analysis tools. A concept label is not a guarantee that a generated answer is correct, that an explanation is complete, or that a model can identify a particular training document. Those questions require separate evidence.

Choose a notebook that matches your question

The upstream examples make the intended analysis more concrete:

Your question Upstream starting point
Which components contribute to a token prediction? Base-model logit contribution notebook
How does the output relate to input tokens? Base-model feature attribution notebook
What happens when a concept is amplified? Base-model concept amplification notebook

For your own evaluation, save the prompt, model revision, generation settings and unmodified output before applying an intervention. Change one setting at a time and compare both the analysis and the generated text. Include examples where your intended concept is absent so you can detect effects outside the target case.

Keep the release and installation instructions together

The reviewed repository and model card show different quick-start forms. Follow one version's instructions and dependency requirements consistently; do not combine snippets from different revisions and call the result a tested tutorial. Review the linked model-weight terms for the release you intend to evaluate, separately from the source-code license.

The CCAPI documentation describes CCAPI's API contract. An upstream Hugging Face model identifier on this page does not establish that CCAPI hosts that model or exposes its internal attribution features through a standard chat endpoint.

We checked the source links and documented interfaces for this revision. We did not load the model, run a GPU benchmark or validate attribution quality. Unsupported performance numbers, personal usage claims and unrelated repository or paper identifiers from the earlier article have been removed.