README.md

metadata

title: Tagline Foundry
emoji: 🏷️
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.12.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Rate your tagline, then let a fine-tuned model beat it.
models:
  - standd/tagline-qwen3p5-4b
  - standd/tagline-quality-setfit
datasets:
  - standd/saas-taglines-rated
  - standd/saas-tagline-distilled

🏷️ Tagline Foundry

Paste a company URL. We render the landing page, grade your current hero tagline against the five marks of a good one, then a fine-tuned Qwen3.5-4B writes sharper alternatives — each scored 0–100 by a SetFit classifier.

Why we built it

This exists to fix our own tagline. We make Hey Lefty — an autonomous research agent that briefs you every morning on any topic you follow (papers, markets, regulators, competitors) and builds a compounding knowledge base while you do your actual work. Our homepage hero just said "Autonomous Research Agents" — a category, not a tagline. Rather than agonize over copy, we trained a model to write a better one. Tagline Foundry is that pipeline, opened up for any site.

How it works

1. Render — lightpanda, an open-source headless browser, loads the page (with JS, so SPAs resolve).
2. Extract — the hero <h1> + og:description become your current tagline and the page context.
3. Critique — the model grades that line ✅/❌ on the five qualities; a SetFit classifier scores it 0–100.
4. Rewrite — a fine-tuned Qwen3.5-4B, distilled from Claude Opus 4.7 as the teacher, drafts a dozen alternatives.
5. Rank — SetFit scores each; the strongest rise to the top.

The Space is free CPU; the GPU work runs on a Modal serverless endpoint that scales to zero, so the first request after idle cold-starts (~60–90s) then runs fast.

Open weights & data

• Generator: standd/tagline-qwen3p5-4b (Apache-2.0)
• Quality scorer: standd/tagline-quality-setfit
• Datasets: saas-taglines-rated · saas-tagline-distilled

Trained via Opus teacher distillation; on held-out companies it out-writes their own taglines ~60% of the time (1-shot, rubric-judged).