---
title: Cardboard Chat
emoji: 🃏
colorFrom: red
colorTo: yellow
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
short_description: Chat with a Pokémon card investment advisor
python_version: "3.12"
startup_duration_timeout: 1h
pinned: false
---

# 🃏 Cardboard Chat

A chat interface for **CardQwen-7B** — a Pokémon card investment advisor
fine-tuned from Qwen2.5-7B-Instruct. Ask it which sets to hold sealed, which
singles to grade, and how to think about market timing.

## ⚠️ Not financial advice

This model produces opinionated reasoning about a collectible market. It is
**not** financial advice. Nothing it says is an offer, a recommendation, or a
guarantee of returns. Card values move; sealed product is illiquid. Do your own
research.

## Usage

1. Tick **"I am 18 or older"** and **"I understand this is not financial advice."**
   (the chat unlocks only after both are checked).
2. Optionally tick the **data-consent** box (see below).
3. Ask away.

## Privacy & data collection

We can collect chat data to improve the model, but only if you explicitly opt
in — the consent checkbox is **off by default** and the chat works without it.

**What we collect, and only when you consent:**
- Your question and the model's reply
- A timestamp
- A random **deletion ID** (shown in the chat UI)
- Your Hugging Face username, *only if* you're logged in (anonymous visitors are
  recorded without any username)

**What we do NOT collect:** IP addresses, browser fingerprints, email addresses,
or any other personal identifier.

**Why:** to build better training data for future versions of CardQwen.

**Where:** a **private** Hugging Face Dataset repo (`gramajo/cardboard-chat-logs`),
accessible only to the maintainer.

**How long:** until you ask us to delete it, or the project ends.

**Your rights:** you may request access to, correction of, or deletion of any
data you've submitted. To do so, **open a Discussion on this Space and quote
your deletion ID** (or your Hugging Face username). We'll remove your rows.

By ticking the consent box you agree to this processing. You can withdraw
consent at any time; it doesn't affect already-usable anonymized model
improvements, but we will delete your raw chats on request.

**Minors:** this Space is intended for adults. The 18+ gate is a hard
requirement — we do not knowingly collect data from anyone under 18.

> Note: Hugging Face infrastructure is hosted in the United States. By opting
> in, you acknowledge your data will be processed there.

## The model

- **Repo:** `gramajo/CardQwen-7B` (private)
- **Base:** Qwen2.5-7B-Instruct
- **System prompt:** the PokéBench advisor prompt baked into the fine-tune
- **Generation:** `temperature=0.35`, `do_sample=True`, `max_new_tokens=256`

## Deploying / maintaining

- **Hardware:** ZeroGPU (`zero-a10g`), set via `--flavor zero-a10g` on creation
  or `hf spaces settings <id> --hardware zero-a10g`.
- **Secret:** `HF_TOKEN` — needs **read** on `gramajo/CardQwen-7B` and **write**
  on `gramajo/cardboard-chat-logs`.
- The logs dataset repo is created automatically (private) on the first
  consented message.
- Logs accumulate in `train.jsonl` — load them directly with
  `datasets.load_dataset("gramajo/cardboard-chat-logs")` for future training.