Merge branch 'main' of https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard into merge_origin

app.py

@@ -42,7 +42,7 @@ from src.tools.plots import (
 )
 
 # Start ephemeral Spaces on PRs (see config in README.md)
-enable_space_ci()
+#enable_space_ci()
 
 def restart_space():
     API.restart_space(repo_id=REPO_ID, token=H4_TOKEN)
@@ -106,13 +106,10 @@ def update_table(
     type_query: list,
     precision_query: str,
     size_query: list,
-    show_deleted: bool,
-    show_merges: bool,
-    show_moe: bool,
-    show_flagged: bool,
+    hide_models: list,
     query: str,
 ):
-    filtered_df = filter_models(hidden_df, type_query, size_query, precision_query, show_deleted, show_merges, show_moe, show_flagged)
+    filtered_df = filter_models(df=hidden_df, type_query=type_query, size_query=size_query, precision_query=precision_query, hide_models=hide_models)
     filtered_df = filter_queries(query, filtered_df)
     df = select_columns(filtered_df, columns)
     return df
@@ -160,21 +157,21 @@ def filter_queries(query: str, filtered_df: pd.DataFrame):
 
 
 def filter_models(
-    df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, show_deleted: bool, show_merges: bool, show_moe:bool, show_flagged: bool
+    df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, hide_models: list
 ) -> pd.DataFrame:
     # Show all models
-    if show_deleted:
-        filtered_df = df
-    else:  # Show only still on the hub models
+    if "Private or deleted" in hide_models:
         filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True]
+    else:
+        filtered_df = df
 
-    if not show_merges:
+    if "Contains a merge/moerge" in hide_models:
         filtered_df = filtered_df[filtered_df[AutoEvalColumn.merged.name] == False]
 
-    if not show_moe:
+    if "MoE" in hide_models:
         filtered_df = filtered_df[filtered_df[AutoEvalColumn.moe.name] == False]
 
-    if not show_flagged:
+    if "Flagged" in hide_models:
         filtered_df = filtered_df[filtered_df[AutoEvalColumn.flagged.name] == False]
 
     type_emoji = [t[0] for t in type_query]
@@ -193,10 +190,7 @@ leaderboard_df = filter_models(
     type_query=[t.to_str(" : ") for t in ModelType], 
     size_query=list(NUMERIC_INTERVALS.keys()), 
     precision_query=[i.value.name for i in Precision],
-    show_deleted=True, 
-    show_merges=False, 
-    show_moe=True,
-    show_flagged=False
+    hide_models=["Contains a merge/moerge", "Flagged"], # "Private or deleted", "Contains a merge/moerge", "Flagged"
 )
 
 demo = gr.Blocks(css=custom_css)
@@ -231,17 +225,11 @@ with demo:
                             interactive=True,
                         )
                     with gr.Row():
-                        deleted_models_visibility = gr.Checkbox(
-                            value=True, label="Show private/deleted models", interactive=True
-                        )
-                        merged_models_visibility = gr.Checkbox(
-                            value=False, label="Show merges", interactive=True
-                        )
-                        moe_models_visibility = gr.Checkbox(
-                            value=True, label="Show MoE", interactive=True
-                        )
-                        flagged_models_visibility = gr.Checkbox(
-                            value=False, label="Show flagged models", interactive=True
+                        hide_models = gr.CheckboxGroup(
+                            label="Hide models",
+                            choices = ["Private or deleted", "Contains a merge/moerge", "Flagged", "MoE"],
+                            value=["Private or deleted", "Contains a merge/moerge", "Flagged"],
+                            interactive=True
                         )
                 with gr.Column(min_width=320):
                     #with gr.Box(elem_id="box-filter"):
@@ -296,10 +284,7 @@ with demo:
                     filter_columns_type,
                     filter_columns_precision,
                     filter_columns_size,
-                    deleted_models_visibility,
-                    merged_models_visibility,
-                    moe_models_visibility,
-                    flagged_models_visibility,
+                    hide_models,
                     search_bar,
                 ],
                 leaderboard_table,
@@ -315,10 +300,7 @@ with demo:
                     filter_columns_type,
                     filter_columns_precision,
                     filter_columns_size,
-                    deleted_models_visibility,
-                    merged_models_visibility,
-                    moe_models_visibility,
-                    flagged_models_visibility,
+                    hide_models,
                     search_bar,
                 ],
                 leaderboard_table,
@@ -326,7 +308,7 @@ with demo:
             # Check query parameter once at startup and update search bar + hidden component
             demo.load(load_query, inputs=[], outputs=[search_bar, hidden_search_bar])
             
-            for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, merged_models_visibility, moe_models_visibility, flagged_models_visibility]:
+            for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, hide_models]:
                 selector.change(
                     update_table,
                     [
@@ -335,10 +317,7 @@ with demo:
                         filter_columns_type,
                         filter_columns_precision,
                         filter_columns_size,
-                        deleted_models_visibility,
-                        merged_models_visibility,
-                        moe_models_visibility,
-                        flagged_models_visibility,
+                        hide_models,
                         search_bar,
                     ],
                     leaderboard_table,
@@ -487,7 +466,7 @@ with demo:
 
 scheduler = BackgroundScheduler()
 scheduler.add_job(restart_space, "interval", seconds=10800) # restarted every 3h
-scheduler.add_job(update_dynamic_files, "cron", minute=00) # launched every hour on the hour
+scheduler.add_job(update_dynamic_files, "cron", minute=30) # launched every hour on the hour
 scheduler.start()
 
 demo.queue(default_concurrency_limit=40).launch()

src/display/about.py

@@ -64,9 +64,8 @@ Side note on the baseline scores:
 ## Icons
 - {ModelType.PT.to_str(" : ")} model: new, base models, trained on a given corpora
 - {ModelType.FT.to_str(" : ")} model: pretrained models finetuned on more data
-Specific fine-tune subcategories (more adapted to chat):
-- {ModelType.IFT.to_str(" : ")} model: instruction fine-tunes, which are model fine-tuned specifically on datasets of task instruction 
-- {ModelType.RL.to_str(" : ")} model: reinforcement fine-tunes, which usually change the model loss a bit with an added policy. 
+- {ModelType.chat.to_str(" : ")} model: chat like fine-tunes, either using IFT (datasets of task instruction), RLHF or DPO (changing the model loss a bit with an added policy), etc
+- {ModelType.merges.to_str(" : ")} model: merges or MoErges, models which have been merged or fused without additional fine-tuning. 
 If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information!
 
 "Flagged" indicates that this model has been flagged by the community, and should probably be ignored! Clicking the link will redirect you to the discussion about the model.

src/display/utils.py

@@ -210,9 +210,9 @@ class ModelDetails:
 
 class ModelType(Enum):
     PT = ModelDetails(name="pretrained", symbol="🟢")
-    FT = ModelDetails(name="fine-tuned", symbol="🔶")
-    IFT = ModelDetails(name="instruction-tuned", symbol="⭕")
-    RL = ModelDetails(name="RL-tuned", symbol="🟦")
+    FT = ModelDetails(name="fine-tuned on domain-specific datasets", symbol="🔶")
+    chat = ModelDetails(name="chat models (RLHF, DPO, IFT, ...)", symbol="💬")
+    merges = ModelDetails(name="base merges and moerges", symbol="🤝")
     Unknown = ModelDetails(name="", symbol="?")
 
     def to_str(self, separator=" "):
@@ -224,10 +224,10 @@ class ModelType(Enum):
             return ModelType.FT
         if "pretrained" in type or "🟢" in type:
             return ModelType.PT
-        if "RL-tuned" in type or "🟦" in type:
-            return ModelType.RL
-        if "instruction-tuned" in type or "⭕" in type:
-            return ModelType.IFT
+        if any([k in type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "⭕", "💬"]]):
+            return ModelType.chat
+        if "merge" in type or "🤝" in type:
+            return ModelType.merges
         return ModelType.Unknown
 
 class WeightType(Enum):

src/leaderboard/filter_models.py

@@ -43,6 +43,13 @@ FLAGGED_MODELS = {
     "dillfrescott/trinity-medium": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
     "udkai/Garrulus": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/526",
     "dfurman/GarrulusMarcoro-7B-v0.1": "https://huggingface.co/dfurman/GarrulusMarcoro-7B-v0.1/discussions/1",
+    "udkai/Turdus": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "eren23/slerp-test-turdus-beagle": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "abideen/NexoNimbus-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "alnrg2arg/test2_3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "nfaheem/Marcoroni-7b-DPO-Merge": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "CultriX/MergeTrix-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "liminerity/Blur-7b-v1.21": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
     # Merges not indicated
     "gagan3012/MetaModelv2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
     "gagan3012/MetaModelv3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
@@ -126,6 +133,6 @@ def remove_forbidden_models(leaderboard_data: list[dict]):
     return leaderboard_data
 
 
-def filter_models(leaderboard_data: list[dict]):
+def filter_models_flags(leaderboard_data: list[dict]):
     leaderboard_data = remove_forbidden_models(leaderboard_data)
     flag_models(leaderboard_data)

src/populate.py

@@ -5,7 +5,7 @@ import pandas as pd
 
 from src.display.formatting import has_no_nan_values, make_requests_clickable_model
 from src.display.utils import AutoEvalColumn, EvalQueueColumn, baseline_row
-from src.leaderboard.filter_models import filter_models
+from src.leaderboard.filter_models import filter_models_flags
 from src.leaderboard.read_evals import get_raw_eval_results
 
 
@@ -13,7 +13,7 @@ def get_leaderboard_df(results_path: str, requests_path: str, dynamic_path: str,
     raw_data = get_raw_eval_results(results_path=results_path, requests_path=requests_path, dynamic_path=dynamic_path)
     all_data_json = [v.to_dict() for v in raw_data]
     all_data_json.append(baseline_row)
-    filter_models(all_data_json)
+    filter_models_flags(all_data_json)
 
     df = pd.DataFrame.from_records(all_data_json)
     df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)

src/scripts/update_all_request_files.py

@@ -3,7 +3,8 @@ from huggingface_hub import ModelCard
 
 import json
 import time
-from src.submission.check_validity import is_model_on_hub, check_model_card
+
+from src.submission.check_validity import is_model_on_hub, check_model_card, get_model_tags
 from src.envs import DYNAMIC_INFO_REPO, DYNAMIC_INFO_PATH, DYNAMIC_INFO_FILE_PATH, API, H4_TOKEN
 
 def update_models(file_path, models):
@@ -28,45 +29,20 @@ def update_models(file_path, models):
             #data['params'] = get_model_size(model_cfg, data['precision'])
             data['license'] = model_cfg.card_data.license if model_cfg.card_data is not None else ""
 
-            # Is the model still on the hub
-            still_on_hub, error, model_config = is_model_on_hub(
-                model_name=model_id, revision=data.get("revision"), trust_remote_code=True, test_tokenizer=False, token=H4_TOKEN
+            # Is the model still on the hub?
+            model_name = model_id
+            if model_cfg.card_data is not None and model_cfg.card_data.base_model is not None:
+                model_name = model_cfg.card_data.base_model # for adapters, we look at the parent model
+            still_on_hub, _, _ = is_model_on_hub(
+                model_name=model_name, revision=data.get("revision"), trust_remote_code=True, test_tokenizer=False, token=H4_TOKEN
             )
             data['still_on_hub'] = still_on_hub
 
             tags = []
 
             if still_on_hub:
-                model = model_id
-                modelcard_OK, error_msg = check_model_card(model)
-                model_card = None
-                if modelcard_OK:
-                    model_card = ModelCard.load(model)
-                
-                is_merge_from_metadata = False
-                is_moe_from_metadata = False
-                is_merge_from_model_card = False
-                is_moe_from_model_card = False
-
-                # Storing the model tags
-                moe_keywords = ["moe", "mixture of experts", "mixtral"]
-                if modelcard_OK:
-                    if model_card.data.tags:
-                        is_merge_from_metadata = "merge" in model_card.data.tags
-                        is_moe_from_metadata = "moe" in model_card.data.tags
-                    merge_keywords = ["mergekit", "merged model", "merge model", "merging"]
-                    # If the model is a merge but not saying it in the metadata, we flag it
-                    is_merge_from_model_card = any(keyword in model_card.text.lower() for keyword in merge_keywords)
-                    if is_merge_from_model_card or is_merge_from_metadata:
-                        tags.append("merge")
-                        if not is_merge_from_metadata:
-                            tags.append("flagged:undisclosed_merge")
-                    is_moe_from_model_card = any(keyword in model_card.text.lower() for keyword in moe_keywords)    
-                is_moe_from_name = "moe" in model.lower().replace("/", "-").replace("_", "-").split("-")
-                if is_moe_from_model_card or is_moe_from_name or is_moe_from_metadata:
-                    tags.append("moe")
-                    if not is_moe_from_metadata:
-                        tags.append("flagged:undisclosed_moe")
+                status, _, model_card = check_model_card(model_id)
+                tags = get_model_tags(model_card, model_id)
 
             data["tags"] = tags
 
@@ -108,3 +84,4 @@ def update_dynamic_files():
         commit_message=f"Daily request file update.",
     )
     print(f"UPDATE_DYNAMIC: pushed to hub")
+

src/submission/check_validity.py

@@ -19,7 +19,7 @@ def check_model_card(repo_id: str) -> tuple[bool, str]:
     try:
         card = ModelCard.load(repo_id)
     except huggingface_hub.utils.EntryNotFoundError:
-        return False, "Please add a model card to your model to explain how you trained/fine-tuned it."
+        return False, "Please add a model card to your model to explain how you trained/fine-tuned it.", None
 
     # Enforce license metadata
     if card.data.license is None:
@@ -27,13 +27,13 @@ def check_model_card(repo_id: str) -> tuple[bool, str]:
             return False, (
                 "License not found. Please add a license to your model card using the `license` metadata or a"
                 " `license_name`/`license_link` pair."
-            )
+            ), None
 
     # Enforce card content
     if len(card.text) < 200:
-        return False, "Please add a description to your model card, it is too short."
+        return False, "Please add a description to your model card, it is too short.", None
 
-    return True, ""
+    return True, "", card
 
 
 def is_model_on_hub(model_name: str, revision: str, token: str = None, trust_remote_code=TRUST_REMOTE_CODE, test_tokenizer=False) -> tuple[bool, str, AutoConfig]:
@@ -133,3 +133,34 @@ def already_submitted_models(requested_models_dir: str) -> set[str]:
                     users_to_submission_dates[organisation].append(info["submitted_time"])
 
     return set(file_names), users_to_submission_dates
+
+def get_model_tags(model_card, model: str):
+    is_merge_from_metadata = False
+    is_moe_from_metadata = False
+    is_merge_from_model_card = False
+    is_moe_from_model_card = False
+
+    # Storing the model tags
+    tags = []
+    moe_keywords = ["moe", "mixture of experts", "mixtral"]
+    if model_card is not None:
+        if model_card.data.tags:
+            is_merge_from_metadata = "merge" in model_card.data.tags
+            is_moe_from_metadata = "moe" in model_card.data.tags
+        merge_keywords = ["merged model", "merge model"]
+        # If the model is a merge but not saying it in the metadata, we flag it
+        is_merge_from_model_card = any(keyword in model_card.text.lower() for keyword in merge_keywords)
+        if is_merge_from_model_card or is_merge_from_metadata:
+            tags.append("merge")
+            if not is_merge_from_metadata:
+                tags.append("flagged:undisclosed_merge")
+        moe_keywords = ["moe", "mixtral"]
+        is_moe_from_model_card = any(keyword in model_card.text.lower() for keyword in moe_keywords)
+    is_moe_from_name = "moe" in model.lower().replace("/", "-").replace("_", "-").split("-")
+    if is_moe_from_model_card or is_moe_from_name or is_moe_from_metadata:
+        tags.append("moe")
+        # We no longer tag undisclosed MoEs
+        #if not is_moe_from_metadata:
+        #    tags.append("flagged:undisclosed_moe")
+
+    return tags

src/submission/submit.py

@@ -13,6 +13,7 @@ from src.submission.check_validity import (
     get_model_size,
     is_model_on_hub,
     user_submission_permission,
+    get_model_tags
 )
 
 REQUESTED_MODELS = None
@@ -97,37 +98,9 @@ def add_new_eval(
         license = None
         #return styled_error("Please select a license for your model")
 
-    modelcard_OK, error_msg = check_model_card(model)
-    model_card = None
-    if modelcard_OK:
-        model_card = ModelCard.load(model)
+    modelcard_OK, error_msg, model_card = check_model_card(model)
     
-    is_merge_from_metadata = False
-    is_moe_from_metadata = False
-    is_merge_from_model_card = False
-    is_moe_from_model_card = False
-
-    # Storing the model tags
-    tags = []
-    moe_keywords = ["moe", "mixture of experts", "mixtral"]
-    if modelcard_OK:
-        if model_card.data.tags:
-            is_merge_from_metadata = "merge" in model_card.data.tags
-            is_moe_from_metadata = "moe" in model_card.data.tags
-        merge_keywords = ["mergekit", "merged model", "merge model", "merging"]
-        # If the model is a merge but not saying it in the metadata, we flag it
-        is_merge_from_model_card = any(keyword in model_card.text.lower() for keyword in merge_keywords)
-        if is_merge_from_model_card or is_merge_from_metadata:
-            tags.append("merge")
-            if not is_merge_from_metadata:
-                tags.append("flagged:undisclosed_merge")
-        is_moe_from_model_card = any(keyword in model_card.text.lower() for keyword in moe_keywords)    
-    is_moe_from_name = "moe" in model.lower().replace("/", "-").replace("_", "-").split("-")
-    if is_moe_from_model_card or is_moe_from_name or is_moe_from_metadata:
-        tags.append("moe")
-        if not is_moe_from_metadata:
-            tags.append("flagged:undisclosed_moe")
-
+    tags = get_model_tags(model_card, model)
 
     # Seems good, creating the eval
     print("Adding new eval")

update_dynamic.py

@@ -0,0 +1,4 @@
+from src.scripts.update_all_request_files import update_dynamic_files
+
+if __name__ == "__main__":
+    update_dynamic_files()