Some rough notes from supervising a social-media chatbot project built on Persian language models.
Placeholder post — replace with real write-up.
A few things worth writing up properly at some point:
- How we evaluated the base models before fine-tuning
- What the Persian tokenization edge cases actually looked like in practice
- The metrics we ended up trusting vs. the ones that looked good but weren't
More detail coming once I turn this into a real post.