[arihanv]

Closes #192

#192 asks for a classification notebook that works with many classes (the reporter has about 400). The text classification notebook linked in the README trims the LM head down to single tokens like "0" to "9", which breaks once labels need more than one token.

This adds Qwen3 (1.7B) Classification, which uses a sequence-classification head instead: FastLanguageModel.from_pretrained(..., num_labels, id2label, label2id), one output per class, trained with cross-entropy. Label names never go through the tokenizer, so the same code works for 77 or 400+ classes.

What the notebook covers

• Banking77: 77 customer-support intents, ~10k train / 3,080 test.
• A short "using your own data" section with class_encode_column, for string labels like the reporter's.
• LoRA via get_peft_model. Unsloth detects the classification model and trains the new score head too.
• Accuracy and macro F1 during and after training, and top-k predictions with probabilities.

Results (free Colab T4, full Run all)

• Accuracy 0.936, macro F1 0.936 on the 3,080 test questions after 2 epochs
• Peak reserved memory about 2.3 GB (15.5% of the T4)

Files

• original_template/Qwen3_(1.7B)-Classification.ipynb, plus the Colab, Kaggle and AMD notebooks and the python_scripts/ and fsdp2_scripts/ copies generated by update_all_notebooks.py
• README: one row each in the Colab, Kaggle and AMD Qwen tables

When I ran update_all_notebooks.py locally, the Hugging Face metadata lookup failed for every model, which reordered most README tables and touched unrelated notebooks. I kept only the files for this notebook and its README rows.