[CI] add back sacrebleu (and document why) (#37700)

* example test

* add back dep

* dev-ci

* dev-ci

Author: gante

setup.py

@@ -163,6 +163,10 @@
     "rjieba",
     "rouge-score!=0.0.7,!=0.0.8,!=0.1,!=0.1.1",
     "ruff==0.11.2",
+    # `sacrebleu` not used in `transformers`. However, it is needed in several tests, when a test calls
+    # `evaluate.load("sacrebleu")`. This metric is used in the examples that we use to test the `Trainer` with, in the
+    # `Trainer` tests (see references to `run_translation.py`).
+    "sacrebleu>=1.4.12,<2.0.0",
     "sacremoses",
     "safetensors>=0.4.3",
     "sagemaker>=2.31.0",
@@ -352,6 +356,7 @@ def run(self):
         "tensorboard",
         "pydantic",
         "sentencepiece",
+        "sacrebleu",  # needed in trainer tests, see references to `run_translation.py`
     )
     + extras["retrieval"]
     + extras["modelcreation"]

src/transformers/dependency_versions_table.py

@@ -69,6 +69,7 @@
     "rjieba": "rjieba",
     "rouge-score": "rouge-score!=0.0.7,!=0.0.8,!=0.1,!=0.1.1",
     "ruff": "ruff==0.11.2",
+    "sacrebleu": "sacrebleu>=1.4.12,<2.0.0",
     "sacremoses": "sacremoses",
     "safetensors": "safetensors>=0.4.3",
     "sagemaker": "sagemaker>=2.31.0",