Implement stage2_model.stage1_context = "sampled" (gitea #41) #66
Assignees
lars (Lars Bogner)
Clear assignees
lars (Lars Bogner)
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: lars/giant#66
Reference in New Issue
Block a user
Delete Branch "fix/issue-41"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Stage 2 was trained on ground-truth stage-1 outcomes but deployed on
sampled ones, and in a rollout that gap compounds over every step of
every track — the same train/inference gap teacher_forcing="scheduled"
already closes within stage 2, just never applied at the stage
boundary. "sampled" was declared in the schema but rejected loudly by
validate_config as unimplemented; this lands the real implementation.
Mirrors the existing scheduled-sampling precedent rather than a hard
switch: new stage2_model.ctx_p_start/ctx_p_end (defaults 1.0 -> 0.0)
linearly ramp P(condition on ground truth) from epoch 0 to the final
epoch, so stage 2 doesn't chase a wildly moving stage-1 target early in
training. Per the plan discussed with the user: the sample is drawn
from stage 1's sampling_model() (EMA weights when present, matching
what inference actually deploys), mixed per example via a Bernoulli
draw (never blended within a row), and validation always uses the
ground truth regardless of the schedule. Fixes a latent bug the same
pattern would otherwise have hit: every sampler in giant/sample.py
flips its model to .eval() with no restore, so sampling from the raw
(non-EMA) stage-1 model mid-step now explicitly restores its .training
flag afterward to avoid silently corrupting stage 1's own training mode
for the rest of the epoch.
validate_config now enforces stage1_context in {"truth", "sampled"},
requires both stages active for "sampled" (nothing to sample from
otherwise), range-checks ctx_p_start/ctx_p_end, and rejects the
ctx_p_start = ctx_p_end = 1.0 configuration as an unadvertised no-op
identical to "truth".
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Stage 2 was trained on ground-truth stage-1 outcomes but deployed on sampled ones, and in a rollout that gap compounds over every step of every track — the same train/inference gap teacher_forcing="scheduled" already closes within stage 2, just never applied at the stage boundary. "sampled" was declared in the schema but rejected loudly by validate_config as unimplemented; this lands the real implementation. Mirrors the existing scheduled-sampling precedent rather than a hard switch: new stage2_model.ctx_p_start/ctx_p_end (defaults 1.0 -> 0.0) linearly ramp P(condition on ground truth) from epoch 0 to the final epoch, so stage 2 doesn't chase a wildly moving stage-1 target early in training. Per the plan discussed with the user: the sample is drawn from stage 1's sampling_model() (EMA weights when present, matching what inference actually deploys), mixed per example via a Bernoulli draw (never blended within a row), and validation always uses the ground truth regardless of the schedule. Fixes a latent bug the same pattern would otherwise have hit: every sampler in giant/sample.py flips its model to .eval() with no restore, so sampling from the raw (non-EMA) stage-1 model mid-step now explicitly restores its .training flag afterward to avoid silently corrupting stage 1's own training mode for the rest of the epoch. validate_config now enforces stage1_context in {"truth", "sampled"}, requires both stages active for "sampled" (nothing to sample from otherwise), range-checks ctx_p_start/ctx_p_end, and rejects the ctx_p_start = ctx_p_end = 1.0 configuration as an unadvertised no-op identical to "truth". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>