Merge branch 'main' into add-custom-model

README.md

@@ -44,12 +44,12 @@ Hub, S3, or locally.
 
 ## 🔑 Key Features
 
-- **Speed**: [Use vllm as backend for fast evals](https://github.com/huggingface/lighteval/wiki/Use-VLLM-as-backend).
-- **Completeness**: [Use the accelerate backend to launch any models hosted on Hugging Face](https://github.com/huggingface/lighteval/wiki/Quicktour#accelerate).
-- **Seamless Storage**: [Save results in S3 or Hugging Face Datasets](https://github.com/huggingface/lighteval/wiki/Saving-and-reading-results).
-- **Python API**: [Simple integration with the Python API](https://github.com/huggingface/lighteval/wiki/Using-the-Python-API).
-- **Custom Tasks**: [Easily add custom tasks](https://github.com/huggingface/lighteval/wiki/Adding-a-Custom-Task).
-- **Versatility**: Tons of [metrics](https://github.com/huggingface/lighteval/wiki/Metric-List) and [tasks](https://github.com/huggingface/lighteval/wiki/Available-Tasks) ready to go.
+- **Speed**: [Use vllm as backend for fast evals](https://huggingface.co/docs/lighteval/use-vllm-as-backend).
+- **Completeness**: [Use the accelerate backend to launch any models hosted on Hugging Face](https://huggingface.co/docs/lighteval/quicktour#accelerate).
+- **Seamless Storage**: [Save results in S3 or Hugging Face Datasets](https://huggingface.co/docs/lighteval/saving-and-reading-results).
+- **Python API**: [Simple integration with the Python API](https://huggingface.co/docs/lighteval/using-the-python-api).
+- **Custom Tasks**: [Easily add custom tasks](https://huggingface.co/docs/lighteval/adding-a-custom-task).
+- **Versatility**: Tons of [metrics](https://huggingface.co/docs/lighteval/metric-list) and [tasks](https://huggingface.co/docs/lighteval/available-tasks) ready to go.
 
 
 ## ⚡️ Installation
@@ -58,7 +58,7 @@ Hub, S3, or locally.
 pip install lighteval
 ```
 
-Lighteval allows for many extras when installing, see [here](https://github.com/huggingface/lighteval/wiki/Installation) for a complete list.
+Lighteval allows for many extras when installing, see [here](https://huggingface.co/docs/lighteval/installation) for a complete list.
 
 If you want to push results to the Hugging Face Hub, add your access token as
 an environment variable:
@@ -106,8 +106,8 @@ Harness and HELM teams for their pioneering work on LLM evaluations.
 ## 🌟 Contributions Welcome 💙💚💛💜🧡
 
 Got ideas? Found a bug? Want to add a
-[task](https://github.com/huggingface/lighteval/wiki/Adding-a-Custom-Task) or
-[metric](https://github.com/huggingface/lighteval/wiki/Adding-a-New-Metric)?
+[task](https://huggingface.co/docs/lighteval/adding-a-custom-task) or
+[metric](https://huggingface.co/docs/lighteval/adding-a-new-metric)?
 Contributions are warmly welcomed!
 
 If you're adding a new feature, please open an issue first.

src/lighteval/models/endpoints/endpoint_model.py

@@ -510,7 +510,7 @@ def greedy_until(
 
         for _, _ in tqdm(
             dataset.splits_start_end_iterator(),
-            total=self.DATASET_SPLITS,
+            total=dataset.num_dataset_splits,
             desc="Splits",
             position=0,
             disable=self.disable_tqdm,
@@ -532,12 +532,15 @@ def greedy_until(
                     responses = asyncio.run(self._async_process_batch_generate(batch))
                 else:
                     responses = self._process_batch_generate(batch)
-                for response in responses:
+                for i, response in enumerate(responses):
                     results.append(
                         GenerativeResponse(
                             result=response.generated_text,
                             logits=[item.logprob for item in response.details.prefill] if returns_logits else None,
-                            truncated_tokens_count=-1,
+                            generated_tokens=[token.id for token in response.details.tokens],
+                            truncated_tokens_count=max(
+                                len(self.tokenizer.encode(batch[i].context)) - self.max_length, 0
+                            ),
                             padded_tokens_count=-1,
                         )
                     )
@@ -556,7 +559,7 @@ def loglikelihood(
 
         for _, _ in tqdm(
             dataset.splits_start_end_iterator(),
-            total=self.DATASET_SPLITS,
+            total=dataset.num_dataset_splits,
             desc="Splits",
             position=0,
             disable=self.disable_tqdm,
@@ -607,7 +610,7 @@ def loglikelihood_rolling(
 
         for _, _ in tqdm(
             dataset.splits_start_end_iterator(),
-            total=self.DATASET_SPLITS,
+            total=dataset.num_dataset_splits,
             desc="Splits",
             position=0,
             disable=self.disable_tqdm,

src/lighteval/models/model_loader.py

@@ -166,7 +166,9 @@ def load_custom_model(config: CustomModelConfig, env_config: EnvConfig):
     return model
 
 
-def load_model_with_inference_endpoints(config: InferenceEndpointModelConfig, env_config: EnvConfig):
+def load_model_with_inference_endpoints(
+    config: Union[InferenceEndpointModelConfig, ServerlessEndpointModelConfig], env_config: EnvConfig
+):
     logger.info("Spin up model using inference endpoint.")
     model = InferenceEndpointModel(config=config, env_config=env_config)
     return model

src/lighteval/models/transformers/transformers_model.py

@@ -880,10 +880,7 @@ def greedy_until(
                     input_ids=tokenized["input_ids"],
                     input_lengths=[len(item == 1) for item in tokenized["attention_mask"]],
                     input_mask=tokenized["attention_mask"],
-                    truncated=[
-                        len(c) - tokenized["input_ids"].shape[1] if len(c) > tokenized["input_ids"].shape[1] else 0
-                        for c in context
-                    ],
+                    truncated=[max(len(c) - tokenized["input_ids"].shape[1], 0) for c in context],
                     padded=[sum(mask == 0) for mask in tokenized["attention_mask"]],
                 )