Add FineWeb2 multilingual gap rerun entrypoint

Author: dlwh

experiments/evals/fineweb2_multilingual.py

@@ -14,6 +14,7 @@
 
 from experiments.defaults import default_tokenize
 from experiments.llama import llama3_tokenizer
+from marin.evaluation.perplexity_gap import RawTextEvaluationDataset, raw_text_dataset
 from marin.execution.executor import executor_main
 from marin.processing.tokenize.data_configs import TokenizerStep
 
@@ -190,6 +191,21 @@ def fineweb2_multilingual_eval_bundle(*, tokenizer: str = llama3_tokenizer) -> d
     )
 
 
+def fineweb2_multilingual_raw_validation_sets(
+    *,
+    configs: Sequence[str] = FINEWEB2_MULTILINGUAL_EVAL_CONFIGS,
+    name_prefix: str = "fineweb2_multilingual",
+) -> dict[str, RawTextEvaluationDataset]:
+    """Return raw FineWeb2 multilingual held-out eval sets for perplexity-gap reports."""
+    return {
+        os.path.join(name_prefix, config): raw_text_dataset(
+            fineweb2_multilingual_parquet_pattern(config, FINEWEB2_EVAL_SPLIT),
+            tags=tuple(fineweb2_multilingual_tags(config)),
+        )
+        for config in configs
+    }
+
+
 if __name__ == "__main__":
     executor_main(
         steps=list(fineweb2_multilingual_eval_bundle().values()),

experiments/exp_model_perplexity_gap_fineweb2_multilingual.py

@@ -0,0 +1,81 @@
+# Copyright The Marin Authors
+# SPDX-License-Identifier: Apache-2.0
+
+from fray.v2.types import ResourceConfig
+
+from experiments.defaults import default_raw_validation_sets
+from experiments.evals.fineweb2_multilingual import fineweb2_multilingual_raw_validation_sets
+from marin.evaluation.perplexity_gap import GapFinderModelConfig, default_model_perplexity_gap
+from marin.execution.executor import executor_main
+
+RESOURCE_CONFIG = ResourceConfig.with_tpu("v5p-8", regions=["us-central1"])
+MAX_DOCS_PER_DATASET = 256
+MAX_DOC_BYTES = 32_768
+
+DATASETS = {
+    **default_raw_validation_sets(),
+    **fineweb2_multilingual_raw_validation_sets(),
+}
+
+MARIN_MODEL = GapFinderModelConfig(
+    checkpoint_path="marin-community/marin-8b-base",
+    checkpoint_is_hf=True,
+    tokenizer="meta-llama/Llama-3.1-8B",
+)
+
+MARIN_VS_LLAMA = default_model_perplexity_gap(
+    name="fineweb2-multilingual-marin-8b-base-vs-llama-3.1-8b-base-doccap256",
+    model_a=MARIN_MODEL,
+    model_b=GapFinderModelConfig(
+        checkpoint_path="meta-llama/Llama-3.1-8B",
+        checkpoint_is_hf=True,
+        tokenizer="meta-llama/Llama-3.1-8B",
+    ),
+    datasets=DATASETS,
+    resource_config=RESOURCE_CONFIG,
+    per_device_batch_size=4,
+    max_eval_length=4096,
+    max_docs_per_dataset=MAX_DOCS_PER_DATASET,
+    max_doc_bytes=MAX_DOC_BYTES,
+    wandb_tags=[
+        "eval=perplexity-gap",
+        "rerun=fineweb2-multilingual",
+        "model_a=marin-community/marin-8b-base",
+        "model_b=meta-llama/Llama-3.1-8B",
+        "dataset_bundle=default_raw_plus_fineweb2_multilingual",
+        "region=us-central1",
+        f"max_docs_per_dataset={MAX_DOCS_PER_DATASET}",
+    ],
+)
+
+MARIN_VS_QWEN3 = default_model_perplexity_gap(
+    name="fineweb2-multilingual-marin-8b-base-vs-qwen3-8b-base-doccap256",
+    model_a=MARIN_MODEL,
+    model_b=GapFinderModelConfig(
+        checkpoint_path="Qwen/Qwen3-8B-Base",
+        checkpoint_is_hf=True,
+        tokenizer="Qwen/Qwen3-8B",
+    ),
+    datasets=DATASETS,
+    resource_config=RESOURCE_CONFIG,
+    per_device_batch_size=4,
+    max_eval_length=4096,
+    max_docs_per_dataset=MAX_DOCS_PER_DATASET,
+    max_doc_bytes=MAX_DOC_BYTES,
+    wandb_tags=[
+        "eval=perplexity-gap",
+        "rerun=fineweb2-multilingual",
+        "model_a=marin-community/marin-8b-base",
+        "model_b=Qwen/Qwen3-8B-Base",
+        "dataset_bundle=default_raw_plus_fineweb2_multilingual",
+        "region=us-central1",
+        f"max_docs_per_dataset={MAX_DOCS_PER_DATASET}",
+    ],
+)
+
+
+if __name__ == "__main__":
+    executor_main(
+        [MARIN_VS_LLAMA, MARIN_VS_QWEN3],
+        description="Run Marin perplexity-gap reports with FineWeb2 multilingual held-out eval sets.",
+    )

tests/evals/test_fineweb2_multilingual.py

@@ -0,0 +1,20 @@
+# Copyright The Marin Authors
+# SPDX-License-Identifier: Apache-2.0
+
+from experiments.evals.fineweb2_multilingual import fineweb2_multilingual_raw_validation_sets
+
+
+def test_fineweb2_multilingual_raw_sets_use_eval_split_in_path_only():
+    datasets = fineweb2_multilingual_raw_validation_sets(configs=("deu_Latn",))
+
+    dataset = datasets["fineweb2_multilingual/deu_Latn"]
+
+    assert isinstance(dataset.input_path, str)
+    assert "/deu_Latn/test/*.parquet" in dataset.input_path
+    assert dataset.split == "validation"
+    assert dataset.tags == (
+        "fineweb2_multilingual",
+        "fineweb2_multilingual/script/Latn",
+        "fineweb2_multilingual/language/deu",
+        "fineweb2_multilingual/top_50_by_rows",
+    )