Sampling past max tokens

Strategies for producing very long outputs beyond a single-call limit, while keeping the result coherent.

• Upstream notebook: misc/sampling_past_max_tokens.ipynb

What to focus on

• Chunked generation plans (outline → sections)
• Continuation prompts that preserve structure
• How to stop safely (avoid runaway generation)

Run locally

make test-notebooks NOTEBOOK=misc/sampling_past_max_tokens.ipynb