Course 3: Post-Training & On-Policy Distillation¶
Paper coverage: Section 5.1 specialist training · Section 5.1.2 OPD · reasoning-effort modes Lab:
posttrain/opd.py· exp11
The pipeline replaced mixed RL entirely [PAPER Section 5.1]¶
Base model
↓ SFT on domain data (per domain: math, coding, agent, instruction-following)
Domain specialists
↓ GRPO reinforcement learning (domain reward models)
RL specialists ← three reasoning-effort modes each
↓ multi-teacher ON-POLICY DISTILLATION
Unified model
More than ten teacher models consolidate into one student. The critical substitution: wherever prior DeepSeek pipelines mixed RL policies, V4 distills instead.
Why reverse KL, why on-policy [PAPER Eq 29]¶
Three design choices carry the weight:
- Student-generated trajectories: teachers score states the student actually visits (on-policy). Off-policy teacher states would supervise a distribution the student never occupies.
- Reverse KL direction \(\mathrm{KL}(\pi_\theta\|\pi_E)\) is mode-seeking: per context the student commits to the relevant expert's behavior instead of averaging all experts into mush.
- Full-vocabulary logits (not sampled-token KL estimates): lower-variance gradients; the token-level shortcut caused instability in prior work.
Teacher scheduling at scale [PAPER Section 5.2.2, see infrastructure course]¶
Weights offloaded to storage; only last-layer hidden states cached; logits rebuilt through the head on demand; samples ordered by teacher index so exactly one head resides in memory.
Reasoning efforts [PAPER Section 5.1.1]¶
Three modes (Non-think / Think High / Think Max) produced by training separate specialists under different RL length penalties and context windows, integrated via <think> scaffolding. Effort is a trained property, not an inference knob.
The lab found the sharp edge [MEASURED, exp11]¶
Our toy consolidation (two synthetic domains, matched budgets) reproduced two failure modes worth knowing before touching OPD:
- Greedy scheduling collapses. Routing each rollout to its most-likely teacher created a feedback ratchet: student drifts toward domain A → A wins scheduling → B decays (loss 2.08 → 3.11). Balanced round-robin scheduling eliminated the collapse.
- Consolidation must clear its overhead. With specialist advantage of only ~0.24 nats, distilled students did not beat matched-budget SFT continuation. OPD's economics require specialists strong enough that logit-level transfer outweighs scheduling noise.
Both findings are consistent with, and explain the necessity of, the paper's heavy teacher-scheduling engineering.