Implement stage2_model.stage1_context = "sampled" (gitea #41)
CI / Lint (ruff check) (push) Successful in 27s
CI / Format (ruff format) (push) Successful in 30s
CI / Sync project version with tag (push) Has been skipped
CI / Type check (ty) (push) Successful in 31s
CI / Tests (push) Successful in 2m26s
CI / Lint (ruff check) (pull_request) Successful in 27s
CI / Format (ruff format) (pull_request) Successful in 27s
CI / Sync project version with tag (pull_request) Has been skipped
CI / Type check (ty) (pull_request) Successful in 27s
CI / Tests (pull_request) Successful in 2m23s

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>
This commit is contained in:
2026-08-17 12:27:13 +02:00
co-authored by Claude Opus 5
parent 09bea2cbff
commit 48faaee79d
7 changed files with 360 additions and 41 deletions
+68 -2
View File
@@ -780,13 +780,79 @@ def test_validate_config_bad_stop_sampling_rejected():
assert "stop_sampling" in str(e)
def test_validate_config_stage1_context_sampled_not_implemented():
def test_validate_config_stage1_context_sampled_accepted_with_both_stages_active():
"""gitea #41: 'sampled' is now implemented, so DEFAULT_CONFIG's
stage1_model/stage2_model.active = true (both) must let it through."""
cfg = _cfg_with(**{"stage2_model.stage1_context": "sampled"})
gconfig.validate_config(cfg) # must not raise
def test_validate_config_bad_stage1_context_rejected():
cfg = _cfg_with(**{"stage2_model.stage1_context": "bogus"})
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "sampled" in str(e)
assert "stage1_context" in str(e)
def test_validate_config_stage1_context_sampled_requires_stage1_active():
cfg = _cfg_with(
**{
"stage2_model.stage1_context": "sampled",
"stage1_model.active": False,
}
)
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "sampled" in str(e) and "stage1_model.active" in str(e)
def test_validate_config_stage1_context_sampled_requires_stage2_active():
cfg = _cfg_with(
**{
"stage2_model.stage1_context": "sampled",
"stage2_model.active": False,
}
)
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "sampled" in str(e) and "stage2_model.active" in str(e)
@pytest.mark.parametrize("key", ["ctx_p_start", "ctx_p_end"])
@pytest.mark.parametrize("value", [-0.1, 1.1])
def test_validate_config_ctx_p_out_of_range_rejected(key, value):
cfg = _cfg_with(
**{
"stage2_model.stage1_context": "sampled",
f"stage2_model.{key}": value,
}
)
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert key in str(e)
def test_validate_config_stage1_context_sampled_always_truth_rejected_as_noop():
cfg = _cfg_with(
**{
"stage2_model.stage1_context": "sampled",
"stage2_model.ctx_p_start": 1.0,
"stage2_model.ctx_p_end": 1.0,
}
)
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "ctx_p_start" in str(e) and "ctx_p_end" in str(e)
def test_validate_config_n_sec_truth_rejected_for_rollout_capable_checkpoint():