fareedkhan-dev--train-llm-from-scratch
wehub-resource-sync
15 行
1.1 KiB
Plaintext
15 行
1.1 KiB
Plaintext
flowchart LR
|
|
Q0([GSM8K prompts]):::data --> Q1["rollout<br/>generate + log-probs"]:::model --> Q2["score<br/>verifier or reward model"]:::rl --> Q3["+ per-token KL to ref"]:::rl --> Q4["GAE<br/>advantages + returns"]:::proc --> Q5["clipped policy + value<br/>(K epochs)"]:::model --> Q6[(ppo.pt)]:::ckpt
|
|
Q5 -. sync old policy .-> Q1
|
|
|
|
classDef data fill:#cdeccd,stroke:#2e7d32,stroke-width:2px,color:#143d1a;
|
|
classDef store fill:#cfeee9,stroke:#1f9e8f,stroke-width:2px,color:#0c3a34;
|
|
classDef proc fill:#cfe3fb,stroke:#1565c0,stroke-width:2px,color:#0d2c52;
|
|
classDef model fill:#ffe8a3,stroke:#d48806,stroke-width:2px,color:#5a3d00;
|
|
classDef rl fill:#ffd9b3,stroke:#e67e22,stroke-width:2px,color:#6b3500;
|
|
classDef loss fill:#ffd4d4,stroke:#c0392b,stroke-width:2px,color:#5c1212;
|
|
classDef ckpt fill:#ececec,stroke:#666666,stroke-width:2px,color:#222222;
|
|
classDef eval fill:#e7d6fb,stroke:#8e44ad,stroke-width:2px,color:#3d1a5a;
|
|
classDef out fill:#e7d6fb,stroke:#7b3fbf,stroke-width:2px,color:#3d1a5a;
|
|
classDef op fill:#ffffff,stroke:#888888,stroke-width:2px,color:#333333;
|