EuroEval 15.4.1__tar.gz → 15.4.2__tar.gz

{euroeval-15.4.1 → euroeval-15.4.2}/.github/ISSUE_TEMPLATE/benchmark_dataset_request.yaml

@@ -2,6 +2,7 @@ name: 📚 Benchmark Dataset Request
 description: Do you think a particular benchmark dataset is missing in EuroEval?
 title: "[BENCHMARK DATASET REQUEST] <dataset-name>"
 labels: "benchmark dataset request"
+type: task
 
 body:
 - type: input
@@ -30,6 +31,7 @@ body:
       - label: Icelandic
       - label: Italian
       - label: Norwegian (Bokmål or Nynorsk)
+      - label: Spanish
       - label: Swedish
   validations:
     required: true

{euroeval-15.4.1 → euroeval-15.4.2}/.github/ISSUE_TEMPLATE/bug.yaml

@@ -1,7 +1,7 @@
 name: 🐛 Bug Report
 description: Have you experienced a bug using the `euroeval` package?
 title: "[BUG] <name-of-bug>"
-labels: bug
+type: bug
 
 body:
 - type: markdown
@@ -46,8 +46,9 @@ body:
       - 3.10.x
       - 3.11.x
       - 3.12.x
+      - 3.13.x
       - Older than 3.10.x
-      - Newer than 3.12.x
+      - Newer than 3.13.x
   validations:
     required: true
 - type: input
@@ -57,6 +58,20 @@ body:
     placeholder: Output of `pip list | grep EuroEval`
   validations:
     required: true
+- type: input
+  attributes:
+    label: Transformers version
+    description: What version of 🤗 transformers are you using?
+    placeholder: Output of `pip list | grep transformers`
+  validations:
+    required: true
+- type: input
+  attributes:
+    label: vLLM version
+    description: What version of vLLM are you using?
+    placeholder: Output of `pip list | grep vllm`
+  validations:
+    required: true
 - type: markdown
   attributes:
     value: >

{euroeval-15.4.1 → euroeval-15.4.2}/.github/ISSUE_TEMPLATE/feature_request.yaml

@@ -1,7 +1,7 @@
 name: 🚀 Feature Request
 description: Is the EuroEval benchmark missing a feature?
 title: "[FEATURE REQUEST] <name-of-feature>"
-labels: enhancement
+type: feature
 
 body:
 - type: textarea
@@ -11,16 +11,6 @@ body:
       A clear and concise description of the feature proposal. Please outline the motivation for the proposal. Is your feature request related to a specific problem? e.g., *"I'm working on X and would like Y to be possible"*.
   validations:
     required: true
-- type: textarea
-  attributes:
-    label: Alternatives
-    description: >
-      A description of any alternative solutions or features you've considered, if any.
-- type: textarea
-  attributes:
-    label: Additional context
-    description: >
-      Add any other context or screenshots about the feature request.
 - type: markdown
   attributes:
     value: >

{euroeval-15.4.1 → euroeval-15.4.2}/.github/ISSUE_TEMPLATE/model_evaluation_request.yaml

@@ -2,6 +2,7 @@ name: 📊 Model Evaluation Request
 description: Would you like to have a particular model included in the leaderboards?
 title: "[MODEL EVALUATION REQUEST] <model-name>"
 labels: "model evaluation request"
+type: task
 
 body:
 - type: input
@@ -10,16 +11,6 @@ body:
     description: What is the Hugging Face model ID?
   validations:
     required: true
-- type: dropdown
-  attributes:
-    label: Model type
-    description: What is the architecture of the model?
-    options:
-      - Decoder model (e.g., GPT)
-      - Encoder model (e.g., BERT)
-      - Sequence-to-sequence model (e.g., T5)
-  validations:
-    required: true
 - type: checkboxes
   attributes:
     label: Evaluation languages
@@ -36,9 +27,29 @@ body:
       - label: Icelandic
       - label: Italian
       - label: Norwegian (Bokmål or Nynorsk)
+      - label: Spanish
       - label: Swedish
   validations:
     required: true
+- type: dropdown
+  attributes:
+    label: Model type
+    description: What is the architecture of the model?
+    options:
+      - Decoder model (e.g., GPT)
+      - Encoder model (e.g., BERT)
+      - Sequence-to-sequence model (e.g., T5)
+  validations:
+    required: true
+- type: dropdown
+  attributes:
+    label: Model size
+    description: What is the size of the model?
+    options:
+      - Small (<=8B parameters)
+      - Large (>8B parameters)
+  validations:
+    required: true
 - type: dropdown
   attributes:
     label: Merged model

{euroeval-15.4.1 → euroeval-15.4.2}/.github/workflows/ci.yaml

@@ -43,7 +43,6 @@ jobs:
       - name: Install uv and set up Python
         uses: astral-sh/setup-uv@v4
         with:
-          enable-cache: true
           python-version: ${{ matrix.python-version }}
 
       - name: Install Dependencies
@@ -75,7 +74,6 @@ jobs:
       - name: Install uv and set up Python
         uses: astral-sh/setup-uv@v4
         with:
-          enable-cache: true
           python-version: ${{ matrix.python-version }}
 
       - name: Install Dependencies

{euroeval-15.4.1 → euroeval-15.4.2}/CHANGELOG.md

@@ -10,6 +10,51 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 
 
 
+## [v15.4.2] - 2025-03-31
+### Added
+- Now added version metadata to results, to easier track which versions of the various
+  dependencies were used when evaluating a model. This currently includes
+  `transformers`, `torch`, `vllm` and `outlines`.
+
+### Changed
+- Changed the name of the German 'mlsum' summarisation dataset to 'mlsum-de', to reflect
+  that it is the German version of the dataset, and to avoid confusion with the Spanish
+  'mlsum-es' dataset.
+
+### Fixed
+- Now uses `fp16` instead of `bf16` when evaluating decoder models on GPUs with CUDA
+  compatibility < 8.0. This was contributed by [@marksverdhei](https://github.com/marksverdhei) ✨
+- Corrected the name of the French sentiment dataset AlloCiné. This was contributed by
+  [@Alkarex](https://github.com/Alkarex) ✨
+- Evaluating a specific model revision did not work for adapter models, as there was a
+  confusion between the revision of the adapter and the revision of the base model. We
+  now use the revision for the adapter and use the latest revision for the base model.
+- In the (very unlikely) scenario that the model's tokeniser has the same first token
+  for two different labels in a text classification task, we now also use the second
+  token to ensure that we determine the correct label. If this is not possible, then we
+  warn the user.
+- Now catches `TypeError` when trying to generate with vLLM, and retries 3 times before
+  giving up on evaluating the dataset.
+- A bug in `transformers` caused models with the `image-text-to-text` pipeline tag to
+  not be detected as generative models. This has been patched now, and will be fixed
+  properly when [this transformers
+  PR](https://github.com/huggingface/transformers/pull/37107) has been merged.
+- Force `vllm` v0.8.0 for now, as the severe degradation in generation output of some
+  models has not been resolved in versions v0.8.2 and v0.8.3.
+- Only accepts the local labels for text classification tasks when evaluating decoder
+  models now, where we before accepted both the local and English labels. The reason is
+  that this caused a confusion mat times when there was a unique local label starting
+  with a particular letter, but a different English label starting with the same letter,
+  causing some models to be evaluated on the wrong label.
+- When fetching the model information from the Hugging Face API we now attempt 3 times,
+  as the API sometimes fails. If it still fails after 3 attempts, we raise the
+  `HuggingFaceHubDown` exception.
+- Now uses `fp16` instead of `bf16` when evaluating decoder models on GPUs with CUDA
+  compatibility < 8.0. This was contributed by [@marksverdhei](https://github.com/marksverdhei) ✨
+- Fixed docs for ScandiQA-da and ScandiQA-sv, where it was incorrectly stated that
+  the splits were made by considering the original train/validation/test splits.
+
+
 ## [v15.4.1] - 2025-03-25
 ### Fixed
 - Disallow `vllm` v0.8.1, as it causes severe degradation in generation output of
@@ -211,7 +256,7 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 
 ### Added
 - Added support for French! 🇫🇷This includes the sentiment classification dataset
-  [Allocine](https://hf.co/datasets/tblard/allocine), the linguistic acceptability
+  [AlloCiné](https://hf.co/datasets/tblard/allocine), the linguistic acceptability
   dataset ScaLA with the [French Universal
   Dependencies](https://github.com/UniversalDependencies/UD_French-GSD), the reading
   comprehension dataset [FQuAD](https://hf.co/datasets/illuin/fquad) (and unofficially

{euroeval-15.4.1 → euroeval-15.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: EuroEval
-Version: 15.4.1
+Version: 15.4.2
 Summary: The robust European language model benchmark.
 Project-URL: Repository, https://github.com/EuroEval/EuroEval
 Project-URL: Issues, https://github.com/EuroEval/EuroEval/issues
@@ -42,6 +42,7 @@ Requires-Dist: more-itertools>=10.5.0
 Requires-Dist: numpy<2.0.0,>=1.23.0
 Requires-Dist: ollama>=0.4.7
 Requires-Dist: pandas>=2.2.0
+Requires-Dist: peft>=0.15.0
 Requires-Dist: protobuf~=3.20.0
 Requires-Dist: pydantic>=2.6.0
 Requires-Dist: pyinfer>=0.0.3
@@ -61,12 +62,12 @@ Requires-Dist: bitsandbytes>=0.43.1; (platform_system == 'Linux') and extra == '
 Requires-Dist: fbgemm-gpu>=1.0.0; (platform_system == 'Linux') and extra == 'all'
 Requires-Dist: gradio>=4.26.0; extra == 'all'
 Requires-Dist: outlines>=0.1.11; extra == 'all'
-Requires-Dist: vllm!=0.8.1,>=0.8.0; (platform_system == 'Linux') and extra == 'all'
+Requires-Dist: vllm==0.8.0; (platform_system == 'Linux') and extra == 'all'
 Provides-Extra: generative
 Requires-Dist: bitsandbytes>=0.43.1; (platform_system == 'Linux') and extra == 'generative'
 Requires-Dist: fbgemm-gpu>=1.0.0; (platform_system == 'Linux') and extra == 'generative'
 Requires-Dist: outlines>=0.1.11; extra == 'generative'
-Requires-Dist: vllm!=0.8.1,>=0.8.0; (platform_system == 'Linux') and extra == 'generative'
+Requires-Dist: vllm==0.8.0; (platform_system == 'Linux') and extra == 'generative'
 Provides-Extra: human-evaluation
 Requires-Dist: gradio>=4.26.0; extra == 'human-evaluation'
 Provides-Extra: test

{euroeval-15.4.1 → euroeval-15.4.2}/docs/datasets/danish.md

@@ -285,11 +285,10 @@ the translated contexts still contained the answer to the question, potentially
 changing the answers slightly.
 
 The original full dataset consists of 6,810 / 500 / 500 samples for training,
-validation and testing, respectively. We use a 1,024 / 256 / 2,048 split for training,
-validation and testing, respectively (so 3,328 samples used in total). All validation
-samples in our version also belong to the original validation set, and all original test
-samples are included in our test set. The remaining 1,548 test samples in our version
-was sampled from the original training set.
+validation and testing, respectively (so 3,328 samples used in total).
+We use a 1,024 / 256 / 2,048 split for training, validation and testing, respectively,
+where the splits are made by randomly sampling from the full dataset without considering
+the original train/validation/test splits.
 
 Here are a few examples from the training split:
 

{euroeval-15.4.1 → euroeval-15.4.2}/docs/datasets/french.md

@@ -7,11 +7,11 @@ information about what these constitute.
 
 ## Sentiment Classification
 
-### Allocine
+### AlloCiné
 
 This dataset was published in [this Github
 repository](https://github.com/TheophileBlard/french-sentiment-analysis-with-bert) and
-features reviews from the French movie review website Allocine. The reviews range from
+features reviews from the French movie review website [AlloCiné](https://www.allocine.fr/). The reviews range from
 0.5 to 5 (inclusive), with steps of 0.5. The negative samples are reviews with a rating
 of at most 2, and the positive ones are reviews with a rating of at least 4. The reviews
 in between were discarded.

{euroeval-15.4.1 → euroeval-15.4.2}/docs/datasets/spanish.md

@@ -475,7 +475,7 @@ $ euroeval --model <model-id> --dataset hellaswag-es
 
 ## Summarization
 
-### MLSum-es-mini
+### MLSum-es
 
 The dataset was published in [this paper](https://aclanthology.org/2020.emnlp-main.647/) and is obtained from online newspapers.
 

{euroeval-15.4.1 → euroeval-15.4.2}/docs/datasets/swedish.md

@@ -231,11 +231,10 @@ the translated contexts still contained the answer to the question, potentially
 changing the answers slightly.
 
 The original full dataset consists of 6,810 / 500 / 500 samples for training,
-validation and testing, respectively. We use a 1,024 / 256 / 2,048 split for training,
-validation and testing, respectively (so 3,328 samples used in total). All validation
-samples in our version also belong to the original validation set, and all original test
-samples are included in our test set. The remaining 1,548 test samples in our version
-was sampled from the original training set.
+validation and testing, respectively (so 3,328 samples used in total).
+We use a 1,024 / 256 / 2,048 split for training, validation and testing, respectively,
+where the splits are made by randomly sampling from the full dataset without considering
+the original train/validation/test splits.
 
 Here are a few examples from the training split:
 

{euroeval-15.4.1 → euroeval-15.4.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "EuroEval"
-version = "15.4.1"
+version = "15.4.2"
 description = "The robust European language model benchmark."
 readme = "README.md"
 authors = [
@@ -39,13 +39,14 @@ dependencies = [
     "setuptools>=75.8.2",
     "demjson3>=3.0.6",
     "ollama>=0.4.7",
+    "peft>=0.15.0",
 ]
 
 [project.optional-dependencies]
 generative = [
     "outlines>=0.1.11",
     "bitsandbytes>=0.43.1; platform_system == 'Linux'",
-    "vllm>=0.8.0,!=0.8.1; platform_system == 'Linux'",
+    "vllm==0.8.0; platform_system == 'Linux'",
     "fbgemm-gpu>=1.0.0; platform_system == 'Linux'",
 ]
 human_evaluation = [
@@ -54,7 +55,7 @@ human_evaluation = [
 all = [
     "outlines>=0.1.11",
     "bitsandbytes>=0.43.1; platform_system == 'Linux'",
-    "vllm>=0.8.0,!=0.8.1; platform_system == 'Linux'",
+    "vllm==0.8.0; platform_system == 'Linux'",
     "fbgemm-gpu>=1.0.0; platform_system == 'Linux'",
     "gradio>=4.26.0",
 ]
@@ -86,7 +87,6 @@ dev-dependencies = [
     "nbstripout>=0.7.1",
     "coverage>=5.5",
     "lxml>=5.1.0",
-    "peft>=0.13.2",
     "mkdocs-material>=9.5.45",
     "mkdocs-include-markdown-plugin>=7.0.1",
     "mkdocs-include-dir-to-nav>=1.2.0",

{euroeval-15.4.1 → euroeval-15.4.2}/src/euroeval/benchmark_modules/hf.py

@@ -20,6 +20,7 @@ from huggingface_hub.utils import (
     HFValidationError,
     LocalTokenNotFoundError,
 )
+from peft import PeftConfig
 from requests.exceptions import RequestException
 from torch import nn
 from transformers import (
@@ -34,6 +35,9 @@ from transformers import (
     Trainer,
 )
 from transformers.modelcard import TASK_MAPPING
+from transformers.models.auto.modeling_auto import (
+    MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES,
+)
 from urllib3.exceptions import RequestError
 
 from ..constants import (
@@ -73,6 +77,7 @@ from ..utils import (
     get_class_by_name,
     get_eos_token,
     internet_connection_available,
+    log_once,
 )
 from .base import BenchmarkModule
 
@@ -727,53 +732,54 @@ def get_model_repo_info(
     # If the model does not exist locally, then we get the model info from the Hugging
     # Face Hub
     if model_info is None:
-        try:
-            model_info = hf_api.model_info(
-                repo_id=model_id, revision=revision, token=token
-            )
-        except (GatedRepoError, LocalTokenNotFoundError) as e:
+        num_attempts = 3
+        for _ in range(num_attempts):
             try:
-                hf_whoami(token=token)
-                logger.warning(
-                    f"Could not access the model {model_id} with the revision "
-                    f"{revision}. The error was {str(e)!r}."
+                model_info = hf_api.model_info(
+                    repo_id=model_id, revision=revision, token=token
                 )
+                break
+            except (GatedRepoError, LocalTokenNotFoundError) as e:
+                try:
+                    hf_whoami(token=token)
+                    logger.warning(
+                        f"Could not access the model {model_id} with the revision "
+                        f"{revision}. The error was {str(e)!r}."
+                    )
+                    return None
+                except LocalTokenNotFoundError:
+                    raise NeedsAdditionalArgument(
+                        cli_argument="--api-key",
+                        script_argument="api_key=<your-api-key>",
+                        run_with_cli=benchmark_config.run_with_cli,
+                    )
+            except (RepositoryNotFoundError, HFValidationError):
                 return None
-            except LocalTokenNotFoundError:
-                raise NeedsAdditionalArgument(
-                    cli_argument="--api-key",
-                    script_argument="api_key=<your-api-key>",
-                    run_with_cli=benchmark_config.run_with_cli,
-                )
-        except (RepositoryNotFoundError, HFValidationError):
-            return None
-        except (OSError, RequestException):
-            if internet_connection_available():
-                raise HuggingFaceHubDown()
-            else:
+            except (OSError, RequestException):
+                if internet_connection_available():
+                    continue
                 raise NoInternetConnection()
+        else:
+            raise HuggingFaceHubDown()
 
     # Get all the Hugging Face repository tags for the model. If the model is an adapter
     # model, then we also get the tags for the base model
     tags = model_info.tags or list()
-    has_base_model_tag = any(
-        tag.startswith("base_model:") and tag.count(":") == 1 for tag in tags
-    )
     base_model_id: str | None = None
-    if has_base_model_tag:
-        has_adapter_config = model_info.siblings is not None and any(
-            sibling.rfilename == "adapter_config.json"
-            for sibling in model_info.siblings
+    has_adapter_config = model_info.siblings is not None and any(
+        sibling.rfilename == "adapter_config.json" for sibling in model_info.siblings
+    )
+    if has_adapter_config:
+        adapter_config = PeftConfig.from_pretrained(model_id, revision=revision)
+        base_model_id = adapter_config.base_model_name_or_path
+        log_once(
+            f"Model {model_id!r} identified as an adapter model, with base model "
+            f"{base_model_id!r}.",
+            level=logging.DEBUG,
         )
-        if has_adapter_config:
-            base_model_id = [
-                tag.split(":")[1]
-                for tag in tags
-                if tag.startswith("base_model:") and tag.count(":") == 1
-            ][0]
+        if base_model_id is not None:
             base_model_info = hf_api.model_info(
                 repo_id=base_model_id,
-                revision=revision,
                 token=benchmark_config.api_key
                 or os.getenv("HUGGINGFACE_API_KEY")
                 or True,
@@ -781,12 +787,18 @@ def get_model_repo_info(
             tags += base_model_info.tags or list()
             tags = list(set(tags))
 
+    # TEMP: This extends the `TASK_MAPPING` dictionary to include the missing
+    # 'image-text-to-text' pipeline tag. This will be added as part of `TASK_MAPPING`
+    # when this PR has been merged in and published:
+    # https://github.com/huggingface/transformers/pull/37107
+    TASK_MAPPING["image-text-to-text"] = MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES
+
     # Get the pipeline tag for the model. If it is not specified, then we determine it
     # by checking the model's architecture as written in the model's Hugging Face config
     pipeline_tag = model_info.pipeline_tag
     if pipeline_tag is None:
         hf_config = load_hf_model_config(
-            model_id=model_id,
+            model_id=base_model_id or model_id,
             num_labels=0,
             id2label=dict(),
             label2id=dict(),
@@ -812,7 +824,6 @@ def get_model_repo_info(
             pipeline_tag = "fill-mask"
 
     if benchmark_config.only_allow_safetensors:
-        # Check if any file ends with .safetensors
         repo_files = hf_api.list_repo_files(repo_id=model_id, revision=revision)
         has_safetensors = any(f.endswith(".safetensors") for f in repo_files)
         if not has_safetensors:
@@ -826,6 +837,26 @@ def get_model_repo_info(
                 )
             raise InvalidModel(msg)
 
+        # Also check base model if we are evaluating an adapter
+        if base_model_id is not None:
+            base_repo_files = hf_api.list_repo_files(repo_id=base_model_id)
+            base_has_safetensors = any(
+                f.endswith(".safetensors") for f in base_repo_files
+            )
+            if not base_has_safetensors:
+                msg = (
+                    f"Base model {base_model_id} does not have safetensors weights "
+                    "available."
+                )
+                if benchmark_config.run_with_cli:
+                    msg += " Skipping since the `--only-allow-safetensors` flag is set."
+                else:
+                    msg += (
+                        " Skipping since the `only_allow_safetensors` argument is set "
+                        "to `True`."
+                    )
+                raise InvalidModel(msg)
+
     return HFModelInfo(
         pipeline_tag=pipeline_tag, tags=tags, adapter_base_model_id=base_model_id
     )

{euroeval-15.4.1 → euroeval-15.4.2}/src/euroeval/benchmark_modules/vllm.py

@@ -30,6 +30,7 @@ from ..constants import (
     REASONING_MAX_TOKENS,
     TASK_GROUPS_USING_LOGPROBS,
     TASKS_USING_JSON,
+    VLLM_BF16_MIN_CUDA_COMPUTE_CAPABILITY,
 )
 from ..data_models import (
     BenchmarkConfig,
@@ -65,6 +66,7 @@ from ..utils import (
     get_bos_token,
     get_end_of_chat_token_ids,
     get_eos_token,
+    get_min_cuda_compute_capability,
     log_once,
     should_prompts_be_stripped,
 )
@@ -145,6 +147,7 @@ class VLLMModel(HuggingFaceEncoderModel):
         if self.model_config.adapter_base_model_id is not None:
             adapter_path = snapshot_download(
                 repo_id=self.model_config.model_id,
+                revision=self.model_config.revision,
                 cache_dir=Path(self.model_config.model_cache_dir),
             )
             self.buffer["lora_request"] = LoRARequest(
@@ -373,12 +376,27 @@ class VLLMModel(HuggingFaceEncoderModel):
 
         # Generate sequences using vLLM
         input_is_a_test = len(prompts) == 1 and len(set(prompts[0])) == 1
-        raw_outputs = self._model.generate(
-            prompts=prompts,
-            sampling_params=sampling_params,
-            use_tqdm=(not input_is_a_test),
-            lora_request=self.buffer.get("lora_request"),
-        )
+        num_attempts = 3
+        for _ in range(num_attempts):
+            try:
+                raw_outputs = self._model.generate(
+                    prompts=prompts,
+                    sampling_params=sampling_params,
+                    use_tqdm=(not input_is_a_test),
+                    lora_request=self.buffer.get("lora_request"),
+                )
+                break
+            except TypeError as e:
+                logger.debug(
+                    f"Encountered error during vLLM generation: {str(e)}. Retrying..."
+                )
+                sleep(1)
+        else:
+            raise InvalidBenchmark(
+                f"Could not generate sequences after {num_attempts} attempts."
+            )
+
+        # Parse the raw model outputs
         completion_ids: list[list[int]] = [
             output.outputs[0].token_ids for output in raw_outputs
         ]
@@ -846,13 +864,16 @@ def load_model_and_tokenizer(
     # Prefer base model ID if the model is an adapter - the adapter will be added on
     # during inference in this case
     model_id = model_config.adapter_base_model_id or model_config.model_id
+    revision = (
+        model_config.revision if model_config.adapter_base_model_id is None else "main"
+    )
 
     hf_model_config = load_hf_model_config(
         model_id=model_id,
         num_labels=0,
         id2label=dict(),
         label2id=dict(),
-        revision=model_config.revision,
+        revision=revision,
         model_cache_dir=model_config.model_cache_dir,
         api_key=benchmark_config.api_key,
         trust_remote_code=benchmark_config.trust_remote_code,
@@ -881,6 +902,23 @@ def load_model_and_tokenizer(
         )
         dtype = torch.float16
 
+    if hf_model_config.torch_dtype == torch.bfloat16:
+        min_cuda_compute_capability = get_min_cuda_compute_capability()
+        required_capability = VLLM_BF16_MIN_CUDA_COMPUTE_CAPABILITY
+
+        if min_cuda_compute_capability is not None:
+            if min_cuda_compute_capability < required_capability:
+                logger.info(
+                    "You are loading a model with "
+                    f"dtype {hf_model_config.torch_dtype}, "
+                    "which vLLM only supports for CUDA devices with"
+                    f"CUDA compute capability >={required_capability}. "
+                    "You are using one or more devices with "
+                    f"compute capability {min_cuda_compute_capability}. "
+                    "Setting dtype to float16 instead."
+                )
+                dtype = torch.float16
+
     if model_config.adapter_base_model_id is not None:
         download_dir = str(Path(model_config.model_cache_dir) / "base_model")
     else:
@@ -916,7 +954,7 @@ def load_model_and_tokenizer(
             max_model_len=min(true_max_model_len, 5_000),
             download_dir=download_dir,
             trust_remote_code=benchmark_config.trust_remote_code,
-            revision=model_config.revision,
+            revision=revision,
             seed=4242,
             distributed_executor_backend=executor_backend,
             tensor_parallel_size=torch.cuda.device_count(),
@@ -994,6 +1032,7 @@ def load_tokenizer(
     Returns:
         The loaded tokenizer.
     """
+    revision = revision if adapter_base_model_id is None else "main"
     config = AutoConfig.from_pretrained(
         adapter_base_model_id or model_id,
         revision=revision,

{euroeval-15.4.1 → euroeval-15.4.2}/src/euroeval/constants.py

@@ -54,3 +54,6 @@ METRIC_ATTRIBUTES_TAKING_UP_MEMORY = ["cached_bertscorer"]
 
 # Hugging Face Hub tags used to classify models as merge models
 MERGE_TAGS = ["merge", "mergekit"]
+
+# The minimum required CUDA compute capability for using bfloat16 in vLLM
+VLLM_BF16_MIN_CUDA_COMPUTE_CAPABILITY = 8.0

{euroeval-15.4.1 → euroeval-15.4.2}/src/euroeval/data_models.py

@@ -1,7 +1,6 @@
 """Data models used in EuroEval."""
 
 import collections.abc as c
-import importlib.metadata
 import json
 import pathlib
 import re
@@ -11,6 +10,8 @@ from dataclasses import dataclass, field
 import pydantic
 import torch
 
+from euroeval.utils import get_package_version
+
 from .enums import Device, InferenceBackend, ModelType, TaskGroup
 from .types import ScoreDict
 
@@ -228,7 +229,11 @@ class BenchmarkResult(pydantic.BaseModel):
     generative_type: str | None
     few_shot: bool
     validation_split: bool
-    euroeval_version: str = importlib.metadata.version("euroeval")
+    euroeval_version: str | None = get_package_version("euroeval")
+    transformers_version: str | None = get_package_version("transformers")
+    torch_version: str | None = get_package_version("torch")
+    vllm_version: str | None = get_package_version("vllm")
+    outlines_version: str | None = get_package_version("outlines")
 
     @classmethod
     def from_dict(cls, config: dict) -> "BenchmarkResult":