Ferr0's picture
ZeroGPU: Qwen2.5-3B + Outlines — schema-conformance demo
a4ab2df verified
|
Raw
History Blame Contribute Delete
2.17 kB
---
title: Structured Output Playground
emoji: 🔒
colorFrom: indigo
colorTo: green
sdk: gradio
sdk_version: "6.19.0"
python_version: "3.12"
app_file: app.py
pinned: true
license: mit
short_description: Lock any LLM's output to a JSON schema
---
# 🔒 Structured Output Playground
**Lock any LLM's output to a JSON schema.** Paste free text, pick (or write) a schema, and a local
model returns structured data that is **guaranteed to conform** — because the decoder is constrained
to the schema at generation time, not asked nicely afterwards.
> The point isn't the model. It's that **schema-conformance is a property of the decoder.**
> Right keys, right types, valid enums — every time.
## The toggle is the demo
There's a **Constraints ON / OFF** switch.
- **ON** — the JSON Schema becomes a grammar; the model can only emit tokens that keep the output
valid *and* conformant. You always get the right shape, the right types, and valid enums.
- **OFF** — the same model just *tries*. A good model often succeeds, but "often" isn't "always":
watch it wrap the JSON in a markdown fence, or — more subtly — return **valid JSON that violates
the schema** (a string where you asked for an integer, a value outside your enum). The **Event**
example is built to show exactly this.
## How it works
- **Model** — [Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct).
- **Inference**`transformers` on **ZeroGPU** (H200).
- **Constraint** — [Outlines](https://github.com/dottxt-ai/outlines) turns the JSON Schema into a
grammar, so only schema-valid token sequences are allowed.
- **Validation** — every output is checked with `jsonschema` so you can *see* conformant vs. broken.
Four presets (contact, product, job posting, event) plus a **Custom** mode where you paste your own
JSON Schema. All example texts are fictional.
## About
Built by **[Ferr0](https://huggingface.co/Ferr0)** — infra-minded AI: local LLM inference,
structured generation & tool-calling, offline RAG, defensive AI security.
More at **[pixelium.win](https://pixelium.win)** · **[GitHub](https://github.com/ferr079)**.
License: MIT.