`1bf4b01`

Use warmup_ratio instead of warmup_steps in DAPT overrides

Authored by mfwolffe <wolffemf@dukes.jmu.edu> 2 weeks ago

Status	File	+	-
M	`src/dlm/train/cpt/runtime.py`	1	5
M	`tests/unit/train/cpt/test_runtime.py`	3	3

src/dlm/train/cpt/runtime.pymodified

      at `learning_rate * min_lr_rate` instead of zero — matching our
      DAPT curve. The warmup-ratio bump to 20% is literature-default for
      domain-adaptive pretraining.
+-
 -    We thread that ratio through `warmup_steps` because current
 -    `TrainingArguments` interprets values `< 1` as ratios and the
 -    dedicated `warmup_ratio` knob is deprecated.
      """
      return {
          "lr_scheduler_type": "cosine_with_min_lr",
 -        "warmup_steps": warmup_ratio,
 +        "warmup_ratio": warmup_ratio,
          "lr_scheduler_kwargs": {"min_lr_rate": floor_ratio},
+     }

tests/unit/train/cpt/test_runtime.pymodified

          ov = dapt_sft_config_overrides()
          assert ov["lr_scheduler_type"] == "cosine_with_min_lr"
 -    def test_warmup_steps_default_20_percent_ratio(self) -> None:
 +    def test_warmup_ratio_default_20_percent(self) -> None:
          ov = dapt_sft_config_overrides()
 -        assert ov["warmup_steps"] == pytest.approx(0.2)
 +        assert ov["warmup_ratio"] == pytest.approx(0.2)
      def test_floor_ratio_default(self) -> None:
          ov = dapt_sft_config_overrides()
      def test_custom_warmup_respected(self) -> None:
          ov = dapt_sft_config_overrides(warmup_ratio=0.3)
 -        assert ov["warmup_steps"] == pytest.approx(0.3)
 +        assert ov["warmup_ratio"] == pytest.approx(0.3)