Allow gating EvaluationTracker HF Hub results; customizability (#2051)

* batch commit

* :Revert "batch commit"

This reverts commit d859d1ca.

* batch commit

* checkout from main

* checkout from main

* checkout from main

* checkout from main

* checkout from main

* cleanup

* cleanup

* cleanup

* cleanup

* cleanup

* cleanup eval results

* cleanup

* add check for gated repo

* fix jsonline issue

* fix

* add try catch when gating the details repo

* add doc

* adds back hub_repo_name

* readds hub repo name

docs/interface.md

@@ -58,12 +58,15 @@ This mode supports a number of command-line arguments, the details of which can
 
 * `--hf_hub_log_args` : Logs evaluation results to Hugging Face Hub. Accepts a string with the arguments separated by commas. Available arguments:
     * `hub_results_org` - organization name on Hugging Face Hub, e.g., `EleutherAI`. If not provided, the results will be pushed to the owner of the Hugging Face token,
-    * `hub_repo_name` - repository name on Hugging Face Hub, e.g., `lm-eval-results`,
+    * `hub_repo_name` - repository name on Hugging Face Hub (deprecated, `details_repo_name` and `results_repo_name` should be used instead), e.g., `lm-eval-results`,
+    * `details_repo_name` - repository name on Hugging Face Hub to store details, e.g., `lm-eval-results`,
+    * `results_repo_name` - repository name on Hugging Face Hub to store results, e.g., `lm-eval-results`,
     * `push_results_to_hub` - whether to push results to Hugging Face Hub, can be `True` or `False`,
     * `push_samples_to_hub` - whether to push samples results to Hugging Face Hub, can be `True` or `False`. Requires `--log_samples` to be set,
     * `public_repo` - whether the repository is public, can be `True` or `False`,
     * `leaderboard_url` - URL to the leaderboard, e.g., `https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard`.
     * `point_of_contact` - Point of contact for the results dataset, e.g., `yourname@example.com`.
+    * `gated` - whether to gate the details dataset, can be `True` or `False`.
 
 ## External Library Usage
 

lm_eval/loggers/evaluation_tracker.py

 import json
+import os
 import re
 import time
 from collections import defaultdict
@@ -14,6 +15,7 @@ from huggingface_hub import (
     HfApi,
     hf_hub_url,
 )
+from huggingface_hub.utils import build_hf_headers, get_session, hf_raise_for_status
 
 from lm_eval.utils import (
     eval_logger,
@@ -112,12 +114,15 @@ class EvaluationTracker:
         output_path: str = None,
         hub_results_org: str = "",
         hub_repo_name: str = "",
+        details_repo_name: str = "",
+        results_repo_name: str = "",
         push_results_to_hub: bool = False,
         push_samples_to_hub: bool = False,
         public_repo: bool = False,
         token: str = "",
         leaderboard_url: str = "",
         point_of_contact: str = "",
+        gated: bool = False,
     ) -> None:
         """
         Creates all the necessary loggers for evaluation tracking.
@@ -126,12 +131,15 @@ class EvaluationTracker:
             output_path (str): Path to save the results. If not provided, the results won't be saved.
             hub_results_org (str): The Hugging Face organization to push the results to. If not provided, the results will be pushed to the owner of the Hugging Face token.
             hub_repo_name (str): The name of the Hugging Face repository to push the results to. If not provided, the results will be pushed to `lm-eval-results`.
+            details_repo_name (str): The name of the Hugging Face repository to push the details to. If not provided, the results will be pushed to `lm-eval-results`.
+            result_repo_name (str): The name of the Hugging Face repository to push the results to. If not provided, the results will not be pushed and will be found in the details_hub_repo.
             push_results_to_hub (bool): Whether to push the results to the Hugging Face hub.
             push_samples_to_hub (bool): Whether to push the samples to the Hugging Face hub.
             public_repo (bool): Whether to push the results to a public or private repository.
             token (str): Token to use when pushing to the Hugging Face hub. This token should have write access to `hub_results_org`.
             leaderboard_url (str): URL to the leaderboard on the Hugging Face hub on the dataset card.
             point_of_contact (str): Contact information on the Hugging Face hub dataset card.
+            gated (bool): Whether to gate the repository.
         """
         self.general_config_tracker = GeneralConfigTracker()
 
@@ -142,6 +150,7 @@ class EvaluationTracker:
         self.leaderboard_url = leaderboard_url
         self.point_of_contact = point_of_contact
         self.api = HfApi(token=token) if token else None
+        self.gated_repo = gated
 
         if not self.api and (push_results_to_hub or push_samples_to_hub):
             raise ValueError(
@@ -159,9 +168,24 @@ class EvaluationTracker:
                 f"hub_results_org was not specified. Results will be pushed to '{hub_results_org}'."
             )
 
-        hub_repo_name = hub_repo_name if hub_repo_name else "lm-eval-results"
-        self.hub_results_repo = f"{hub_results_org}/{hub_repo_name}"
-        self.hub_results_repo_private = f"{hub_results_org}/{hub_repo_name}-private"
+        if hub_repo_name == "":
+            details_repo_name = (
+                details_repo_name if details_repo_name != "" else "lm-eval-results"
+            )
+            results_repo_name = (
+                results_repo_name if results_repo_name != "" else details_repo_name
+            )
+        else:
+            details_repo_name = hub_repo_name
+            results_repo_name = hub_repo_name
+            eval_logger.warning(
+                "hub_repo_name was specified. Both details and results will be pushed to the same repository. Using hub_repo_name is no longer recommended, details_repo_name and results_repo_name should be used instead."
+            )
+
+        self.details_repo = f"{hub_results_org}/{details_repo_name}"
+        self.details_repo_private = f"{hub_results_org}/{details_repo_name}-private"
+        self.results_repo = f"{hub_results_org}/{results_repo_name}"
+        self.results_repo_private = f"{hub_results_org}/{results_repo_name}-private"
 
     def save_results_aggregated(
         self,
@@ -211,9 +235,9 @@ class EvaluationTracker:
 
                 if self.api and self.push_results_to_hub:
                     repo_id = (
-                        self.hub_results_repo
+                        self.results_repo
                         if self.public_repo
-                        else self.hub_results_repo_private
+                        else self.results_repo_private
                     )
                     self.api.create_repo(
                         repo_id=repo_id,
@@ -221,10 +245,15 @@ class EvaluationTracker:
                         private=not self.public_repo,
                         exist_ok=True,
                     )
-                    self.api.upload_folder(
+                    self.api.upload_file(
                         repo_id=repo_id,
-                        folder_path=str(path),
-                        path_in_repo=self.general_config_tracker.model_name_sanitized,
+                        path_or_fileobj=str(
+                            path.joinpath(f"results_{self.date_id}.json")
+                        ),
+                        path_in_repo=os.path.join(
+                            self.general_config_tracker.model_name,
+                            f"results_{self.date_id}.json",
+                        ),
                         repo_type="dataset",
                         commit_message=f"Adding aggregated results for {self.general_config_tracker.model_name}",
                     )
@@ -278,6 +307,7 @@ class EvaluationTracker:
                     sample["resps"] = sanitize_list(sample["resps"])
                     sample["filtered_resps"] = sanitize_list(sample["filtered_resps"])
                     sample["arguments"] = arguments
+                    sample["target"] = str(sample["target"])
 
                     sample_dump = (
                         json.dumps(
@@ -293,9 +323,9 @@ class EvaluationTracker:
 
                 if self.api and self.push_samples_to_hub:
                     repo_id = (
-                        self.hub_results_repo
+                        self.details_repo
                         if self.public_repo
-                        else self.hub_results_repo_private
+                        else self.details_repo_private
                     )
                     self.api.create_repo(
                         repo_id=repo_id,
@@ -303,6 +333,18 @@ class EvaluationTracker:
                         private=not self.public_repo,
                         exist_ok=True,
                     )
+                    try:
+                        if self.gated_repo:
+                            headers = build_hf_headers()
+                            r = get_session().put(
+                                url=f"https://huggingface.co/api/datasets/{repo_id}/settings",
+                                headers=headers,
+                                json={"gated": "auto"},
+                            )
+                            hf_raise_for_status(r)
+                    except Exception as e:
+                        eval_logger.warning("Could not gate the repository")
+                        eval_logger.info(repr(e))
                     self.api.upload_folder(
                         repo_id=repo_id,
                         folder_path=str(path),
@@ -327,9 +369,7 @@ class EvaluationTracker:
         """
 
         eval_logger.info("Recreating metadata card")
-        repo_id = (
-            self.hub_results_repo if self.public_repo else self.hub_results_repo_private
-        )
+        repo_id = self.details_repo if self.public_repo else self.details_repo_private
 
         files_in_repo = self.api.list_repo_files(repo_id=repo_id, repo_type="dataset")
         results_files = get_results_filenames(files_in_repo)
@@ -360,7 +400,10 @@ class EvaluationTracker:
                 results_datetime,
             )
             latest_task_results_datetime[samples_key] = latest_datetime
-            latest_task_results_datetime[results_key] = latest_datetime
+            latest_task_results_datetime[results_key] = max(
+                latest_task_results_datetime[results_key],
+                latest_datetime,
+            )
 
         # Create metadata card
         card_metadata = MetadataConfigs()
@@ -377,14 +420,15 @@ class EvaluationTracker:
             sanitized_last_eval_date_results = re.sub(
                 r"[^\w\.]", "_", latest_task_results_datetime[config_name]
             )
-            # Ensure that all results files are listed in the metadata card
-            current_results = card_metadata.get(config_name, {"data_files": []})
-            current_results["data_files"].append(
-                {"split": eval_date_sanitized, "path": [str(results_filename)]}
-            )
-            card_metadata[config_name] = current_results
-            # If the results file is the newest, update the "latest" field in the metadata card
+
             if eval_date_sanitized == sanitized_last_eval_date_results:
+                # Ensure that all results files are listed in the metadata card
+                current_results = card_metadata.get(config_name, {"data_files": []})
+                current_results["data_files"].append(
+                    {"split": eval_date_sanitized, "path": [str(results_filename)]}
+                )
+                card_metadata[config_name] = current_results
+                # If the results file is the newest, update the "latest" field in the metadata card
                 card_metadata[config_name]["data_files"].append(
                     {"split": "latest", "path": [str(results_filename)]}
                 )
@@ -403,65 +447,20 @@ class EvaluationTracker:
             sanitized_last_eval_date_results = re.sub(
                 r"[^\w\.]", "_", latest_task_results_datetime[config_name]
             )
-            # Ensure that all sample results files are listed in the metadata card
-            current_details_for_task = card_metadata.get(
-                config_name, {"data_files": []}
-            )
-            current_details_for_task["data_files"].append(
-                {"split": eval_date_sanitized, "path": [str(results_filename)]}
-            )
-            card_metadata[config_name] = current_details_for_task
-            # If the samples results file is the newest, update the "latest" field in the metadata card
             if eval_date_sanitized == sanitized_last_eval_date_results:
+                # Ensure that all sample results files are listed in the metadata card
+                current_details_for_task = card_metadata.get(
+                    config_name, {"data_files": []}
+                )
+                current_details_for_task["data_files"].append(
+                    {"split": eval_date_sanitized, "path": [str(results_filename)]}
+                )
+                card_metadata[config_name] = current_details_for_task
+                # If the samples results file is the newest, update the "latest" field in the metadata card
                 card_metadata[config_name]["data_files"].append(
                     {"split": "latest", "path": [str(results_filename)]}
                 )
 
-            # Special case for MMLU with a single split covering it all
-            # We add another config with all MMLU splits results together for easy inspection
-            SPECIAL_TASKS = ["mmlu", "gpqa", "minerva_math"]
-            for special_task in SPECIAL_TASKS:
-                if special_task in config_name:
-                    special_task = f"{model_name}__{special_task}"
-                    former_entry = card_metadata.get(special_task, {"data_files": []})
-
-                    former_split = [
-                        (i, entry)
-                        for i, entry in enumerate(former_entry["data_files"])
-                        if entry.get("split", None) == eval_date_sanitized
-                    ]
-
-                    if len(former_split) == 0:
-                        former_entry["data_files"].append(
-                            {
-                                "split": eval_date_sanitized,
-                                "path": [str(results_filename)],
-                            }
-                        )
-                    else:
-                        split_index, _ = former_split[0]
-                        former_entry["data_files"][split_index]["path"].append(
-                            str(results_filename)
-                        )
-
-                    if eval_date_sanitized == sanitized_last_eval_date_results:
-                        latest_split = [
-                            (i, entry)
-                            for i, entry in enumerate(former_entry["data_files"])
-                            if entry.get("split", None) == "latest"
-                        ]
-                        if len(latest_split) == 0:
-                            former_entry["data_files"].append(
-                                {"split": "latest", "path": [str(results_filename)]}
-                            )
-                        else:
-                            latest_index, _ = latest_split[0]
-                            former_entry["data_files"][latest_index]["path"].append(
-                                str(results_filename)
-                            )
-
-                    card_metadata[special_task] = former_entry
-
         # Get latest results and extract info to update metadata card examples
         latest_datetime = max(latest_task_results_datetime.values())
         latest_model_name = max(

lm_eval/loggers/utils.py

@@ -118,15 +118,15 @@ def add_tokenizer_info(storage: Dict[str, Any], lm):
             tokenizer_info = {
                 "tokenizer_pad_token": [
                     lm.tokenizer.pad_token,
-                    lm.tokenizer.pad_token_id,
+                    str(lm.tokenizer.pad_token_id),
                 ],
                 "tokenizer_eos_token": [
                     lm.tokenizer.eos_token,
-                    lm.tokenizer.eos_token_id,
+                    str(lm.tokenizer.eos_token_id),
                 ],
                 "tokenizer_bos_token": [
                     lm.tokenizer.bos_token,
-                    lm.tokenizer.bos_token_id,
+                    str(lm.tokenizer.bos_token_id),
                 ],
                 "eot_token_id": getattr(lm, "eot_token_id", None),
                 "max_length": getattr(lm, "max_length", None),

lm_eval/utils.py

@@ -163,7 +163,7 @@ def get_file_datetime(filename: str) -> str:
     """
     Given the results and sample results filenames, extracts and returns the datetime.
     """
-    return filename[filename.rfind("_") + 1 :].replace(".json", "")
+    return filename[filename.rfind("_") + 1 :].replace(".jsonl", "")
 
 
 def sanitize_model_name(model_name: str) -> str: