Pre-baked Colab notebooks for open-source chatbots

What it does

This repository is a curated collection of Jupyter notebooks that launch the oobabooga/text-generation-webui Gradio interface inside a Google Colab session. Each notebook targets a specific open-source model—Vicuna, Llama 2, Code Llama, Mistral, Pygmalion, and others—often using GPTQ 4-bit quantized weights to fit inside a free Colab session. Think of it as a set of pre-wired power adapters for running large language models in a cloud sandbox.

The interesting bit

The project’s value is entirely in the curation and the memory math. The maintainer has matched each model to a quantization level—4-bit, 8-bit, or 16-bit—so the notebooks actually execute on Colab hardware without the usual manual configuration. It is essentially glue code, but it saves the configuration slog that usually kills an impulse to test a new model.

Key highlights

• Covers a wide zoo of models, from instruction-tuned chatbots (WizardLM, Koala) to code models (Code Llama) and role-play fine-tunes (Pygmalion).
• Mostly relies on GPTQ 4-bit quantization to squeeze 7B and 13B parameter models into limited GPU memory.
• Wraps the existing text-generation-webui project rather than reinventing the UI.
• Includes both research-licensed and commercially permissible models (Llama 2).

Caveats

• Several notebooks are marked 🚦 WIP 🚦, including some Llama-2 variants, so stability is mixed.
• The README explicitly flags licensing confusion around original LLaMA and whether Colab Pro counts as commercial use.
• An IndexError: list index out of range can crop up; the fix requires manually setting the model’s instruction template in the UI.

Verdict

Worth a bookmark if you want to quickly benchmark or compare open-source chatbots without maintaining a local GPU rig. Skip it if you are looking for a production API, a novel training pipeline, or a standalone application—this is purely a launcher collection.