EuroEval 16.3.0__py3-none-any.whl → 16.4.0__py3-none-any.whl

euroeval/benchmark_modules/litellm.py

@@ -6,7 +6,7 @@ import json
 import logging
 import re
 import typing as t
-from functools import cache, cached_property, partial
+from functools import cached_property, partial
 from time import sleep
 
 import litellm
@@ -36,8 +36,8 @@ from litellm.utils import supports_reasoning, supports_response_schema
 from pydantic import conlist, create_model
 from requests.exceptions import RequestException
 from tqdm.asyncio import tqdm as tqdm_async
-from tqdm.auto import tqdm
 
+from ..caching_utils import cache_arguments
 from ..constants import (
     JSON_STRIP_CHARACTERS,
     LITELLM_CLASSIFICATION_OUTPUT_KEY,
@@ -70,6 +70,7 @@ from ..generation_utils import (
     extract_few_shot_examples,
     raise_if_wrong_params,
 )
+from ..logging_utils import get_pbar, log, log_once
 from ..task_group_utils import (
     question_answering,
     sequence_classification,
@@ -83,7 +84,6 @@ from ..utils import (
     add_semaphore_and_catch_exception,
     create_model_cache_dir,
     get_hf_token,
-    log_once,
     safe_run,
     split_model_id,
 )
@@ -95,8 +95,6 @@ if t.TYPE_CHECKING:
     from litellm.types.utils import ModelResponse
     from transformers.trainer import Trainer
 
-logger = logging.getLogger("euroeval")
-
 
 VOCAB_SIZE_MAPPING = {
     # OpenAI models
@@ -133,6 +131,7 @@ MODEL_MAX_LENGTH_MAPPING = {
     r"gpt-4.1.*": 1_047_576,
     # Anthropic models
     r"(anthropic/)?claude-[1-9](-[1-9])?-(opus|sonnet|haiku)-[0-9]{8}": 200_000,
+    r"(anthropic/)?claude-(opus|sonnet|haiku)-[1-9](-[1-9])?-[0-9]{8}": 200_000,
     # Gemini models
     r"(gemini/)?gemini-1\.5-flash.*": 1_048_576,
     r"(gemini/)?gemini-1\.5-pro.*": 2_097_152,
@@ -367,10 +366,11 @@ class LiteLLMModel(BenchmarkModule):
                 (batch_indices[idx], model_inputs[batch_indices[idx]])
                 for idx, _ in failures
             ]
-            logger.debug(
+            log(
                 f"Attempt {attempt + 1:,}/{num_attempts:,}: retrying "
                 f"{len(inputs_to_run):,} failed message(s). Here is the first error: "
-                f"{failures[0][1]}."
+                f"{failures[0][1]}.",
+                level=logging.DEBUG,
             )
 
             # Attempt to handle the exceptions, to improve the chance of getting
@@ -422,14 +422,19 @@ class LiteLLMModel(BenchmarkModule):
             "'stop' is not supported with this model",
             "'$.stop' is invalid",
         ]
+        stop_pattern = re.compile(r"does not support parameters: \[.*'stop'.*\]")
         logprobs_messages = [
             "you are not allowed to request logprobs",
             "you've reached the maximum number of requests with logprobs",
             "logprobs is not supported",
             "logprobs is not enabled",
+            "Invalid value at 'generation_config.response_logprobs' (TYPE_BOOL)",
         ]
-        top_logprobs_messages = ["got an unexpected keyword argument 'top_logprobs'"]
         logprobs_pattern = re.compile(
+            r"does not support parameters: \[.*'logprobs'.*\]"
+        )
+        top_logprobs_messages = ["got an unexpected keyword argument 'top_logprobs'"]
+        top_logprobs_pattern = re.compile(
             r"does not support parameters: \[.*'top_logprobs'.*\]"
         )
         max_completion_tokens_pattern = re.compile(
@@ -438,6 +443,7 @@ class LiteLLMModel(BenchmarkModule):
         temperature_messages = [
             "'temperature' is not supported with this model.",
             "temperature is not supported with this model",
+            r"does not support parameters: \[.*'temperature'.*\]",
         ]
         temperature_must_be_one_messages = [
             "`temperature` may only be set to 1",
@@ -454,10 +460,14 @@ class LiteLLMModel(BenchmarkModule):
         requires_thinking_disabled_messages = ["thinking.type: Field required"]
         seed_pattern = re.compile(r"does not support parameters: \[.*'seed'.*\]")
         response_format_messages = [
-            "got an unexpected keyword argument 'response_format'"
+            "got an unexpected keyword argument 'response_format'",
+            "the model returned empty outputs",
         ]
 
-        if any(msg.lower() in error_msg for msg in stop_messages):
+        if (
+            any(msg.lower() in error_msg for msg in stop_messages)
+            or stop_pattern.search(string=error_msg) is not None
+        ):
             log_once(
                 f"The model {model_id!r} does not support "
                 "stop sequences, so disabling them.",
@@ -467,7 +477,7 @@ class LiteLLMModel(BenchmarkModule):
             return generation_kwargs
         elif (
             any(msg.lower() in error_msg for msg in logprobs_messages)
-            or logprobs_pattern.search(string=error_msg)
+            or logprobs_pattern.search(string=error_msg) is not None
             # Special case for Vertex AI models, since they have strict rate
             # limits on using logprobs. They also have a cap of 5 logprobs, but
             # we ignore this since the rate limiting makes it unusable anyway.
@@ -477,10 +487,15 @@ class LiteLLMModel(BenchmarkModule):
                 f"The model {model_id!r} does not support logprobs, so disabling it.",
                 level=logging.DEBUG,
             )
+            self.buffer["first_label_token_mapping"] = False
             generation_kwargs.pop("logprobs", None)
             generation_kwargs.pop("top_logprobs", None)
+            generation_kwargs.pop("response_format", None)
             return generation_kwargs
-        elif any(msg.lower() in error_msg for msg in top_logprobs_messages):
+        elif (
+            any(msg.lower() in error_msg for msg in top_logprobs_messages)
+            or top_logprobs_pattern.search(string=error_msg) is not None
+        ):
             log_once(
                 f"The model {model_id!r} does not support the `top_logprobs` argument, "
                 "so moving the value to `logprobs`.",
@@ -597,9 +612,10 @@ class LiteLLMModel(BenchmarkModule):
         elif isinstance(
             error, (Timeout, ServiceUnavailableError, InternalServerError, SystemError)
         ):
-            logger.debug(
+            log(
                 f"Service temporarily unavailable. The error message was: {error}. "
-                "Retrying in 10 seconds..."
+                "Retrying in 10 seconds...",
+                level=logging.DEBUG,
             )
             sleep(10)
             return generation_kwargs
@@ -629,10 +645,32 @@ class LiteLLMModel(BenchmarkModule):
             ) from error
 
         if isinstance(error, RateLimitError):
-            raise InvalidModel(
+            log(
                 f"You have encountered your rate limit for model {model_id!r}. "
-                "Skipping."
-            ) from error
+                "Retrying in 10 seconds...",
+                level=logging.DEBUG,
+            )
+            sleep(10)
+            return generation_kwargs
+
+        if (
+            isinstance(error, BadRequestError)
+            and (
+                retry_match := re.search(
+                    pattern=r"\bretry in ([0-9]+(.[0-9]+)?) ?(s|seconds)\b",
+                    string=error_msg,
+                )
+            )
+            is not None
+        ):
+            retry_seconds = float(retry_match.group(1))
+            log(
+                f"Bad request error encountered. Retrying in {retry_seconds:.1f} "
+                "seconds...",
+                level=logging.DEBUG,
+            )
+            sleep(retry_seconds)
+            return generation_kwargs
 
         if isinstance(error, AuthenticationError):
             raise NeedsAdditionalArgument(
@@ -711,7 +749,19 @@ class LiteLLMModel(BenchmarkModule):
                 for input_ in inputs
                 if isinstance(input_, list)
             ]
-        responses = await tqdm_async.gather(*requests, leave=False)
+        responses = await tqdm_async.gather(
+            *requests, colour="yellow", ascii="—▰", leave=False
+        )
+
+        # If the outputs are empty, convert them to exceptions
+        if all(
+            not isinstance(response, Exception)
+            and response.choices[0].message.content == "{}"
+            for response in responses
+        ):
+            responses = [ValueError("The model returned empty outputs.")] * len(
+                responses
+            )
 
         # Separate the successful responses from the failed ones
         successes = [
@@ -731,7 +781,9 @@ class LiteLLMModel(BenchmarkModule):
                 try:
                     request.close()
                 except RuntimeError as e:
-                    logger.debug(f"RuntimeError during request.close(): {e}")
+                    log(
+                        f"RuntimeError during request.close(): {e}", level=logging.DEBUG
+                    )
 
         return successes, failures
 
@@ -756,10 +808,11 @@ class LiteLLMModel(BenchmarkModule):
         for model_response in model_responses:
             if not model_response.choices:
                 sequences.append("")
-                logger.warning(
+                log(
                     f"The model {model_id!r} did not end up "
                     "generating any text. This is likely because the model ran "
-                    "out of tokens while reasoning. Returning an empty string."
+                    "out of tokens while reasoning. Returning an empty string.",
+                    level=logging.WARNING,
                 )
                 continue
 
@@ -847,11 +900,12 @@ class LiteLLMModel(BenchmarkModule):
                 scores.append(logprobs_list)
 
         if not sequences:
-            logger.warning(
+            log(
                 "No sequences were generated by the model "
                 f"{model_id!r}. This may be due to the "
                 "model running out of tokens or an issue with the input data. "
-                "Returning an empty GenerativeModelOutput."
+                "Returning an empty GenerativeModelOutput.",
+                level=logging.WARNING,
             )
             return GenerativeModelOutput(sequences=[], scores=None)
 
@@ -1130,6 +1184,7 @@ class LiteLLMModel(BenchmarkModule):
                 return partial(
                     sequence_classification.extract_labels_from_generation,
                     dataset_config=self.dataset_config,
+                    model_config=self.model_config,
                     first_label_token_mapping=self.buffer["first_label_token_mapping"],
                 )
             case TaskGroup.TEXT_TO_TEXT:
@@ -1205,17 +1260,19 @@ class LiteLLMModel(BenchmarkModule):
                 ServiceUnavailableError,
                 InternalServerError,
             ) as e:
-                logger.debug(
+                log(
                     f"Service temporarily unavailable. The error message was: {e}. "
-                    "Retrying in 10 seconds..."
+                    "Retrying in 10 seconds...",
+                    level=logging.DEBUG,
                 )
                 sleep(10)
             except APIError as e:
                 if "'503 Service Unavailable" not in str(e):
                     raise e
-                logger.warning(
+                log(
                     f"Failed to check if model {model_id!r} exists. Retrying in 10 "
-                    "seconds..."
+                    "seconds...",
+                    level=logging.WARNING,
                 )
                 sleep(10)
             except (BadRequestError, NotFoundError):
@@ -1228,21 +1285,25 @@ class LiteLLMModel(BenchmarkModule):
                     case 0:
                         pass
                     case 1:
-                        logger.warning(
+                        log(
                             f"Could not find the model ID {model_id!r}. Did you mean "
-                            f"{candidate_models[0]!r}?"
+                            f"{candidate_models[0]!r}?",
+                            level=logging.WARNING,
                         )
                     case _:
                         candidate_models_str = "', '".join(candidate_models)
-                        logger.warning(
+                        log(
                             f"Could not find the model ID {model_id!r}. Did you mean "
-                            f"any of the following model IDs: '{candidate_models_str}'?"
+                            "any of the following model IDs: "
+                            f"'{candidate_models_str}'?",
+                            level=logging.WARNING,
                         )
                 return False
         else:
-            logger.error(
+            log(
                 f"Failed to check if model {model_id!r} exists after {num_attempts} "
-                "attempts. Assuming it does not exist."
+                "attempts. Assuming it does not exist.",
+                level=logging.ERROR,
             )
             return False
 
@@ -1275,7 +1336,8 @@ class LiteLLMModel(BenchmarkModule):
                 "that the revision is actually the parameter and set the revision "
                 "to 'main'. In the future, use the new '#' syntax to specify the "
                 f"parameter (in this case, this would be {proper_model_id!r}), as this "
-                "will be an error in future versions of EuroEval."
+                "will be an error in future versions of EuroEval.",
+                level=logging.WARNING,
             )
             model_id_components.param = model_id_components.revision
             model_id_components.revision = "main"
@@ -1363,7 +1425,7 @@ class LiteLLMModel(BenchmarkModule):
 
         return dataset
 
-    @cache
+    @cache_arguments()
     def get_generation_kwargs(self, dataset_config: DatasetConfig) -> dict[str, t.Any]:
         """Get the generation arguments for the model.
 
@@ -1571,7 +1633,8 @@ def try_download_ollama_model(model_id: str) -> bool:
                         f"The model {model_id!r} cannot be found on Ollama, but the "
                         f"model {model_id_with_prefix} *was* found, so we would "
                         "recommend you cancelling this run and trying the evaluation "
-                        "with that model ID instead."
+                        "with that model ID instead.",
+                        level=logging.WARNING,
                     )
                     return False
                 except ollama.ResponseError as inner_e:
@@ -1589,11 +1652,8 @@ def try_download_ollama_model(model_id: str) -> bool:
                 ) from e
 
         # Download the model
-        with tqdm(
-            desc=f"Downloading {ollama_model_id}",
-            unit_scale=True,
-            unit="B",
-            leave=False,
+        with get_pbar(
+            desc=f"Downloading {ollama_model_id}", unit_scale=True, unit="B"
         ) as pbar:
             for status in response:
                 if status.total is not None: