Increase grad_accum to 8, add gc.collect and empty_cache before training

Manually pre-tokenize with apply_chat_template, cap seq length to 512 for OOM fix

Cap turns to 8, require assistant end, 5k conversations for OOM fix

Reduce to 1 epoch, 10k conversations for faster training on larger GPU

Remove system role, verify alternating user/assistant for Gemma compatibility

Add system prompt, cap dataset, filter non-user-first conversations

Filter None text values to fix TypeError in join

Fix alternating roles and use dtype instead of torch_dtype

Merge consecutive same-role turns to fix alternating user/assistant constraint

Remove max_seq_length from SFTTrainer (not supported in current TRL)

Fix max_seq_length placement and torch_dtype deprecation

Upload raw training script as train.py