README.md

metadata

title: Robe Iniesta Persona Chat
emoji: 🎸
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 5.12.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Spanish persona chat — Qwen 7B + LoRA, under 32B
tags:
  - build-small-hackathon
  - zerogpu
  - spanish
  - persona
  - lora
  - qwen2
models:
  - Qwen/Qwen2.5-7B-Instruct
  - kabesaml/robe-iniesta-lora

Robe Iniesta — Persona Chat

Fine-tuned Qwen2.5-7B-Instruct (7B parameters) with a LoRA adapter trained on public interview transcripts. SFT + DPO pipeline. Well under the 32B parameter cap.

⚠️ Fan project. This is not Robe Iniesta. It may hallucinate biographical facts.

Live Space: https://huggingface.co/spaces/build-small-hackathon/robe-iniesta

Hackathon submission assets

Requirement	Status	Link
Model ≤32B	✅ Qwen2.5-7B (7B)	base model
Gradio app in Build Small org	✅	This Space
Demo video	⬜ TODO	REPLACE: https://youtu.be/YOUR_DEMO_ID
Social media post	⬜ TODO	REPLACE: https://x.com/kabesaml/status/YOUR_POST_ID
ZeroGPU	✅	@spaces.GPU(duration=120) in app.py

Demo video

Demo video: https://youtu.be/YOUR_DEMO_ID

1–3 min walkthrough showing:

1. Opening the Space and waiting for the first GPU response (cold start)
2. Asking "¿Qué significa Extremoduro para ti?"
3. A follow-up question to show multi-turn chat
4. Brief note that it's a style simulation, not the real person

Upload to YouTube (unlisted is fine) or attach video to an X post, then replace the URL above.

Social media post

Social post: https://x.com/kabesaml/status/YOUR_POST_ID

Example post text:

Built a Spanish persona chatbot for the @huggingface Build Small hackathon 🎸

Qwen2.5-7B (7B) + LoRA, SFT → DPO on public interview transcripts.
Runs on ZeroGPU. Under the 32B cap.

Try it: https://huggingface.co/spaces/build-small-hackathon/robe-iniesta

#BuildSmall #HuggingFace #Qwen #LoRA

Post on X or LinkedIn, then paste the post URL above.

Model & training

Component	Detail
Base model	Qwen/Qwen2.5-7B-Instruct — 7B params
Adapter	kabesaml/robe-iniesta-lora
Training	LoRA SFT → DPO preference tuning on Modal A100
Quantization	4-bit NF4 at inference (fits ZeroGPU)
Source repo	https://github.com/kabesaml/robellm (update if public)

Space hardware

Set in Space Settings → Hardware → ZeroGPU.

The chat handler uses @spaces.GPU(duration=120). First request after idle may take ~30–60s (model load).

Local development

From the robellm repo:

uv sync --extra dev
uv pip install -e ".[train]" spaces
uv run python space/app.py