taylorsatulataylorsatulaCommunityยท1 Agent Skills Included

TeaLeaves

Visualize and measure how models attend to prompts

Captures how a language model distributes attention across every region of a prompt at every layer. Generates heatmaps, cooking curves, animated layer sweeps, and logit lens projections for any HuggingFace model. Replaces guesswork-based prompt tuning with measurable before-and-after comparisons and multi-seed stability checks. Runs analysis on remote GPU boxes and produces comparison reports locally.
npx skills add taylorsatula/TeaLeaves --all -g -y
Available:

Instructs the agent to act as a mechanistic interpretability researcher that interprets attention data, validates prompt changes against cooking curves and logit lens metrics, and follows the pipeline's technical constraints when running experiments.

All Skills in This Repository (1)

Pure Emerald Level Indicators

Frequently Asked Questions

FAQPage Schema
How to install TeaLeaves?โ–ผ

Run `npx skills add taylorsatula/TeaLeaves --all -g -y` in your terminal to install the skill globally.

What does TeaLeaves measure?โ–ผ

It captures attention weights at every layer of a model's forward pass, showing exactly which regions of your prompt the model focuses on and when it commits to output tokens.

How does TeaLeaves improve prompt engineering?โ–ผ

Instead of eyeballing outputs, you compare attention heatmaps and cooking curves before and after each prompt change to see measurably what improved or degraded.

Which models does TeaLeaves support?โ–ผ

It works with any HuggingFace transformer model, auto-discovering the architecture, layer count, and attention modules from the model config.

Do I need a GPU to use TeaLeaves?โ–ผ

Yes for the attention capture step, which runs on a remote GPU box such as Vast.ai; rendering and analysis of results run locally on any machine.

Related Repositories in Data & Analytics

View All in Data & Analyticsโ†’