Generation

How models produce text — sampling, temperature, decoding

🔴 MUST

🧠 Understand👁️ Visualize✏️ Calculate💻 Code🔬 Experiment🗣️ Explain

8 subtopics — click to study

Logits 🔴 Probabilities 🔴 Greedy Decoding 🟡 Sampling 🔴 Temperature 🔴 Top-K 🟡 Top-P 🟡 Stop Conditions 🟡

🧪

RASL Pull · Sprint S-06 ⭐ Transition — Parallel Learning

You're reading this because RASL is blocked on: when should the recovery_policy stop retrying — what does a principled stopping condition look like?

Stop conditions and EOS token handling in generation are the conceptual blueprint for RASL's safe-stop policy. "Stop generating when done" = "stop retrying when evidence stops improving." Same policy, different scale. Return to reliable-agent-lab/agent/recovery_policy.py and continue building. Read only what unblocks you.

🔨 Lab: M-07 → infer.py with custom stop conditions 🔁 Stop conditions = RASL safe-stop policy 📋 View full parallel plan →

← Back to Roadmap