added varied data for training

[rudra-singh1]

changes:

• added varied training data (more diverse + 500 more tickets)
• trained for 2000 iterations instead of 1000 (same 3B model though)
• appended "do not include empty field tasks. only include properties that were actually found and redacted" to system_prompt in redact.py

goal: with these changes, model shouldn't overfit to training data anymore and adaptively redact PII versus trying to fill in a template