Add extended task for LiveCodeBench codegeneration (#548)

```shell
lighteval vllm \
    "pretrained=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,dtype=float16,data_parallel_size=4,max_model_length=32768,gpu_memory_utilisation=0.8,generation_parameters={temperature: 0.7}" \
    "extended|lcb:codegeneration|0|0" \
    --use-chat-template
```

Author: plaguss

docs/source/use-vllm-as-backend.mdx

@@ -58,3 +58,32 @@ model: # Model specific parameters
 > [!WARNING]
 > In the case of OOM issues, you might need to reduce the context size of the
 > model as well as reduce the `gpu_memory_utilization` parameter.
+
+
+## Dynamically changing the metric configuration
+
+For special kinds of metrics like `Pass@K` or LiveCodeBench's `codegen` metric, you may need to pass specific values like the number of
+generations. This can be done in the `yaml` file in the following way:
+
+```yaml
+model: # Model specific parameters
+  base_params:
+    model_args: "pretrained=HuggingFaceTB/SmolLM-1.7B,revision=main,dtype=bfloat16" # Model args that you would pass in the command line
+  generation: # Generation specific parameters
+    temperature: 0.3
+    repetition_penalty: 1.0
+    frequency_penalty: 0.0
+    presence_penalty: 0.0
+    seed: 42
+    top_k: 0
+    min_p: 0.0
+    top_p: 0.9
+  metric_options: # Optional metric arguments
+    codegen_pass@1:16:
+      num_samples: 16
+```
+
+An optional key `metric_options` can be passed in the yaml file,
+using the name of the metric or metrics, as defined in the `Metric.metric_name`.
+In this case, the `codegen_pass@1:16` metric defined in our tasks will have the `num_samples` updated to 16,
+independently of  the number defined by default.

src/lighteval/main_vllm.py

@@ -134,9 +134,11 @@ def vllm(
         with open(model_args, "r") as f:
             config = yaml.safe_load(f)["model"]
         model_args = config["base_params"]["model_args"]
+        metric_options = config.get("metric_options", {})
         generation_parameters = GenerationParameters.from_dict(config)
     else:
-        generation_parameters = GenerationParameters()
+        generation_parameters = GenerationParameters.from_model_args(model_args)
+        metric_options = {}
 
     model_args_dict: dict = {k.split("=")[0]: k.split("=")[1] if "=" in k else True for k in model_args.split(",")}
     model_config = VLLMModelConfig(**model_args_dict, generation_parameters=generation_parameters)
@@ -146,6 +148,7 @@ def vllm(
         pipeline_parameters=pipeline_params,
         evaluation_tracker=evaluation_tracker,
         model_config=model_config,
+        metric_options=metric_options,
     )
 
     pipeline.evaluate()

src/lighteval/models/model_input.py

@@ -59,6 +59,34 @@ def from_dict(cls, config_dict: dict):
         """
         return GenerationParameters(**config_dict.get("generation", {}))
 
+    @classmethod
+    def from_model_args(cls, model_args: str):
+        """Creates a GenerationParameters object from a model_args string.
+
+        It's used when the model_args are passed as a string in the command line.
+        The generation parameters must follow the following format (at any place in the string):
+        "generation_parameters={key1:value1,key2=value2}"
+
+        Args:
+            model_args (str): A string like the following:
+                "pretrained=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,dtype=float16,max_model_length=32768,generation={temperature:0.7,top_p:5}"
+        """
+
+        def parse_model_args(model_args):
+            import json
+            import re
+
+            pattern = re.compile(r"(\w+)=(\{.*\}|[^,]+)")
+            matches = pattern.findall(model_args)
+            for key, value in matches:
+                key = key.strip()
+                if key == "generation_parameters":
+                    gen_params = re.sub(r"(\w+):", r'"\1":', value)
+                    return json.loads(gen_params)
+
+        params: dict = parse_model_args(model_args) or {}
+        return GenerationParameters(**params)
+
     def to_litellm_dict(self) -> dict:
         """Selects relevant generation and sampling parameters for litellm models.
         Doc: https://docs.litellm.ai/docs/completion/input#input-params-1

src/lighteval/pipeline.py

@@ -132,6 +132,7 @@ def __init__(
         evaluation_tracker: EvaluationTracker,
         model_config=None,
         model=None,
+        metric_options=None,
     ):
         if not (model or model_config):
             raise ValueError("Must provide either a model or model config when creating a pipeline.")
@@ -145,6 +146,7 @@ def __init__(
 
         self.model_config = model_config
         self.evaluation_tracker = evaluation_tracker
+        self._metric_options = metric_options or {}
         self.accelerator, self.parallel_context = self._init_parallelism_manager()
         self.model = self._init_model(model_config, model)
 
@@ -209,6 +211,10 @@ def _init_tasks_and_requests(self, tasks: str):
             )
             task_names_list, fewshots_dict = taskinfo_selector(tasks, registry)
             task_dict = registry.get_task_dict(task_names_list)
+            # If there are metric_options defined from the yaml file,
+            # review if they have to be updated.
+            if self._metric_options:
+                self._update_num_samples(task_dict)
             LightevalTask.load_datasets(list(task_dict.values()), self.pipeline_parameters.dataset_loading_processes)
 
             self.evaluation_tracker.task_config_logger.log(task_dict)
@@ -230,6 +236,19 @@ def _init_tasks_and_requests(self, tasks: str):
             self.requests = requests
             self.docs = docs
 
+    def _update_num_samples(self, task_dict: dict[str, LightevalTask]):
+        """Helper function to update the num_samples of a given metric via the yaml file.
+        As it has to be done at the metric level, it's better to update the value per metric.
+        It will add a num_samples to the already defined metrics' num_samples if defined in the yaml file.
+        As later when constructing the requests the max is taken over the num_samples, this is valid.
+        """
+        for _, task in task_dict.items():
+            for metric in task.metrics:
+                if metric_data := self._metric_options.get(metric.metric_name, None):
+                    num_samples = metric_data.get("num_samples", None)
+                    if num_samples:
+                        task.num_samples = [num_samples]
+
     def _init_random_seeds(self):
         logger.info("--- INIT SEEDS ---")
         random.seed(1234)

src/lighteval/tasks/extended/__init__.py

@@ -25,12 +25,13 @@
 
 if can_load_extended_tasks():
     import lighteval.tasks.extended.ifeval.main as ifeval
+    import lighteval.tasks.extended.lcb.main as lcb
     import lighteval.tasks.extended.mix_eval.main as mix_eval
     import lighteval.tasks.extended.mt_bench.main as mt_bench
     import lighteval.tasks.extended.olympiade_bench.main as olympiad_bench
     import lighteval.tasks.extended.tiny_benchmarks.main as tiny_benchmarks
 
-    AVAILABLE_EXTENDED_TASKS_MODULES = [ifeval, tiny_benchmarks, mt_bench, mix_eval, olympiad_bench]
+    AVAILABLE_EXTENDED_TASKS_MODULES = [ifeval, tiny_benchmarks, mt_bench, mix_eval, olympiad_bench, lcb]
 
 else:
     AVAILABLE_EXTENDED_TASKS_MODULES = []