guidelabs / steerling

What is steerling?

Steerling is a Python package delivering an 8B-parameter causal diffusion language model focused on interpretability. It handles non-autoregressive text generation via confidence-based unmasking, decomposes predictions into 33k known concepts for attribution, and lets you steer outputs by intervening on concept activations. Load it with `pip install steerling` and `SteerlingGenerator.from_pretrained("guidelabs/steerling-8b")` for GPU inference needing ~18GB VRAM.

Why is it gaining traction?

It combines diffusion-style parallel generation with causal interpretability, letting users attribute logits to interpretable concepts and steer like a "steering motor" for precise control – rare in 8B models. Stands out for interpretable causal inference, concept decomposition, and embeddings (hidden, known, unknown), appealing to devs chasing causally interpretable meta-analysis or representation learning without black-box opacity. Early GitHub interest from interpretable ML and causal AI crowds.

Who should use this?

Researchers in interpretable causal inference analyzing wearable sensors, distributional data, or biological pathways via language models. ML engineers building steerable generation for treatment effect estimation or neural causal models. Teams needing quick attribution in text pipelines, like interpretable AI prototypes.

Verdict

Solid inference start for interpretable causal diffusion LMs, with clean pip API and blog docs, but 93 stars and 1.0% credibility signal alpha maturity – no fine-tuning or full evals yet. Worth a spin for interpretability experiments; monitor for instruction-tuned releases.

(198 words)