Today my Hindi NER fine-tune broke mid-train ride

Today was one of those garbage days where I thought I could squeeze model work between a bus ride to Palolem and a late afternoon walk on the beach, so at 11:13 AM IST I SSHed into my AWS g4dn.xlarge spot (T4) from my phone hotspot and started a fine-tune of bert-base-multilingual-cased on my tiny TouristHIN-ner-v1 dataset made from TripAdvisor reviews, using transformers 4.30.2, PyTorch 1.13.1, CUDA 11.7, Python 3.9, seq_len 128, batch_size 32, lr 3e-5, seed 42; training started fine but after the first eval the loss plunged to 0.01 and eval F1 crashed to 0.02 with the model predicting all O tags, and because my hotspot kept dropping when the bus entered a tunnel I only saw partial logs so I assumed data noise or class imbalance and kept going; later on the beach with actual Wi-Fi I pulled the full logs and found no fatal exception, just a nasty label misalignment caused by my quick edit to convert_conll.py at 09:02 AM where I tried to strip punctuation and accidentally removed blank lines so sentence boundaries merged, and on top of that I tokenized without is_split_into_words=True so word-to-token mapping was never used and labels got shifted by the [CLS] token while my label_ids array ended up shorter than input_ids but Trainer quietly continued training; debugging sucked because the training run had already overwritten the checkpoint every hour, I had an old checkpoint at step 1200 that looked fine and a latest one at 5400 that memorized nothing, so I rewrote the preprocessing on my laptop to keep empty line markers, fixed convert_conll.py line 78 where I was popping labels incorrectly, changed tokenizer call to tokenizer(sent_tokens, is_split_into_words=True, truncation=True, max_length=128) and remapped labels with word_ids, then relaunched a clean run at 16:45 IST and after one epoch eval F1 went up to 0.62 on the same seed which proved the problem was alignment not model size; lesson learned the hard way while trying to work between bus stops and chai stalls, next time I'm not touching preprocessing when I'm off a proper keyboard and stable net.