A tax-assessor chat helper concatenates User: lines by hand, and the instruct model replies like a raw completion. What should the code apply?
Select an answer to reveal the explanation.
Short Explanation
Hand-built User: lines make an instruct model reply like a raw completion. Apply the checkpoint’s chat or special-token template so the roles wrap the way the model expects. NCCL greetings, image packers, and per-turn Triton restarts will not restore that format.
Full Explanation
Instruct checkpoints expect a chat or special-token template applied in code, not a homemade User: concatenation. That wrapper is a Domain 2 implementation step, not a Domain 1 prompt-concept lecture. NCCL process groups, image frame packers, and per-turn Triton restarts do not restore the model’s expected role format. Apply the checkpoint template and the completion style should match the instruct model.