🛠️ Lab 2 — Toolchain Setup and Open Model License Audit (50 pts)
Goal: Install the complete 2026 fine-tuning toolchain in a free Google Colab notebook, verify your environment by loading a small open model and running a forward pass, then audit five candidate open models on the Hugging Face Hub for license fitness in a hypothetical commercial deployment. Steps:
- Open a new notebook at colab.research.google.com and set the runtime to a free GPU (Runtime → Change runtime type → T4 GPU).
- In the first cell, install the toolchain:
pip install -U transformers peft trl bitsandbytes accelerate datasets. - Run
import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0)) and confirm your notebook detects a GPU. Paste this output into your submission. - Load a small open model with
from_pretrained — use either Qwen/Qwen2.5-0.5B-Instruct or an equivalent small SmolLM2 checkpoint from Hugging Face. Load both the tokenizer and the model. - Tokenize a short test prompt (e.g., "Explain fine-tuning in one sentence.") and run a forward pass through the model. Confirm you get output back with no errors, and print the output shape or generated text as proof.
- Browse the Hugging Face Hub (huggingface.co/models) and use the license filter in the left sidebar to identify candidate models. Select 5 open models suitable for a hypothetical commercial chatbot deployment (include at least Qwen 3, one Llama version, one Mistral model, and two of your choosing).
- For each model, open its model card on Hugging Face and record: the exact license name, whether it is OSI-approved, any usage caps or field-of-use restrictions, and whether you would greenlight it for commercial fine-tuning and deployment. Flag any clause that could block a "serve the model via our own API" business model.
- Build a comparison table (markdown or spreadsheet) with columns: Model \| License \| OSI-Approved? \| Usage Cap/Restrictions \| Commercial-Deployment Verdict.
- Write 3–4 sentences summarizing which model you'd choose for the hypothetical deployment and why, citing the specific license clause that drove your decision.
Deliverables: Submit your Colab notebook (.ipynb, shared link or file upload) showing the GPU check and successful forward pass, plus your 5-model license comparison table with the written summary, either as a section in the notebook or a separate document upload.