Hybrid AGPT Experiments
Purpose: test whether deterministic whole-tree AGPT is best used as the broad
coverage phase, followed by a smaller stochastic refinement phase using the same
attention f_theta model.
The motivation is narrow: standalone stochastic traversal-stop training with context-only ancestors now beats the clean pd1 baseline on held-out metrics, but only after much more wall time. A hybrid run asks whether pd1 can do the cheap global pass and stochastic sampling can do the final local refinement.
Baselines
Clean pd1 baseline:
rnd/stochastic-agpt/20260611T160456-d64l2-depth16-pd1-100ep
fixed-token PPL: 4.7929
rolling byte PPL: 5.3359
train wall: 582.1s
optimizer steps: ~6,500
Best standalone stochastic row so far:
rnd/lightning-agpt/20260612T204543-d64l2-depth16-lightning-random-desc-u40000-r4-contextonly-co
fixed-token PPL: 4.7637
rolling byte PPL: 5.1657
train wall: 1999.2s
optimizer steps: 160,000
Important correction: that run's filename/config said random-descendants, but a
stale bin/agpt_experiment dropped lightning.anchor_mode from
resolved_config.yml. The trainer banner shows it actually ran
anchor_mode=traversal-stop. Literal uniform random-descendant sampling was
tested later and mostly selected tiny subtrees.
Results
| run | actual sampler | LR | trained query passes | train wall | fixed PPL | rolling byte PPL | note |
|---|---|---|---|---|---|---|---|
20260612T213817... |
traversal-stop | 3e-4 |
397.9M | 266.2s | 5.5280 | 5.7552 | stale orchestrator dropped intended random-descendants; too destructive |
20260612T215213... |
random-descendants | 3e-4 |
0.9M | 21.3s | 5.6295 | 6.2570 | literal uniform nodes, mostly tiny subtrees |
20260612T215441... |
traversal-stop | 3e-5 |
397.9M | 276.7s | 5.4332 | 5.7765 | gentler but still destructive |
Interpretation
The simple hybrid did not work. Starting from the pd1-100 checkpoint and then
resetting Adam for a stochastic pass damages the solution even at 3e-5.
The failure is still useful:
- Stochastic refinement is not plug-and-play with a trained pd1 checkpoint.
- Reset Adam state is probably part of the problem; the stochastic gradients may also be miscalibrated relative to the pd1 optimum.
- Literal uniform random-descendant sampling is not the policy that produced the good stochastic row. It gives too many tiny subtrees unless we add a size/coverage policy.
- If this hybrid line continues, the next credible version needs either optimizer-state continuity, a much smaller trust-region style update, or a sampler constrained to the same large-unit regime as traversal-stop.