---
license: mit
pipeline_tag: text-generation
sdk: gradio
app_file: app.py
hardware: zerogpu
tags:
- qwen
- lora
- peft
- gradio
- zerogpu
- benjamin-franklin
- historical-character
- local-llm
---

# Qwen (Ben) Franklin Model Zoo

This repository is a portfolio of custom Benjamin Franklin LoRA adapters trained on Qwen-family base models. It includes 1.7B, 4B, and 7B experiments covering persona style, identity persistence, English-only dialogue, tool-call cleanup, factual-history repair, natural conversation, and coherence testing.

Open the static model-zoo index:

- `index.html`

Try the interactive Hugging Face / Gradio demo:

- `app.py`

The demo is designed to run as a Hugging Face Space with ZeroGPU (`hardware: zerogpu`). It lazily loads a selected public Qwen base model plus one local LoRA adapter from `adapters/`, then generates a Franklin-style response.

Main folders:

- `adapters/` — copied LoRA adapter artifacts, one folder per model
- `model_cards/` — per-adapter model cards with strengths, weaknesses, data mix, benchmark notes, and compute requirements
- `benchmarks/franklin_coherence/` — copied benchmark JSON/HTML artifacts used by the index and model cards
- `manifest.json` — machine-readable model inventory

Notes:

- The LoRA adapters are intended to be loaded with their listed public base model IDs such as `unsloth/Qwen2.5-7B-Instruct-bnb-4bit`, `unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit`, or `unsloth/Qwen3-1.7B-unsloth-bnb-4bit`.
- The 7B family is the largest Qwen-family Franklin LoRA family proven trainable on an 8GB RTX 3070-class GPU in this project.
- The benchmark scores are lightweight offline evaluation scores, not universal quality claims.
- For factual historical reliability, especially the Craven Street bones / William Hewson story, retrieval or prompt-context is still recommended.