GGUF export only script

Stage 2 only: tool prediction on personality LoRA, 3 epochs, batch=2

Fix: disable eval (OOM at step 40 — model uses 77/80GB). Train loss converging fine.

Fix: DataCollatorForSeq2Seq for proper label padding, removes eval crash at step 40

Fix: extract text tokenizer from VLM Processor to avoid image_utils error

Fix: pre-tokenize with input_ids+labels instead of relying on SFTTrainer auto-tokenization

Fix: use messages format directly instead of text column for SFTTrainer 5.x compat

Fix: remove_unused_columns=False for transformers 5.x, warmup_steps instead of warmup_ratio

Upload train_hf_jobs.py with huggingface_hub

initial commit