Add assemble_corpus.py with combined text+audio multimodal support for Unsloth"

Fix: Remove TeleAntiFraud from text pipeline — it's audio-only (Phase 2). Add clear Phase 1/2 notes."

Add assemble_corpus.py — merges 3 scam datasets into unified training corpus

Fix: Unsloth requires formatting_func — use pre-tokenized text column instead of messages"

Fix: dtype must be None (not string 'auto'). Unsloth only accepts None/torch.float16/torch.bfloat16"

Fix: dtype must be None or torch.dtype, not string. Kaggle T4 supports fp16 auto-detected.

Add 8GB VRAM version of training script (RTX 5060 8GB / Kaggle T4)

Add updated Unsloth SFT script for Gemma 4 E2B scam fine-tuning

Upload eval_zero_shot_cpu.py

Upload README.md

Upload eval_zero_shot_cpu.py

Upload format_dataset.py

Upload README.md

Upload train_sft_unsloth.py

Upload eval_zero_shot.py

initial commit