update benchmark

.gitattributes

@@ -25,6 +25,7 @@
 *.safetensors filter=lfs diff=lfs merge=lfs -text
 saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
 *.tflite filter=lfs diff=lfs merge=lfs -text
 *.tgz filter=lfs diff=lfs merge=lfs -text
 *.wasm filter=lfs diff=lfs merge=lfs -text
@@ -32,4 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
-scale-hf-logo.png filter=lfs diff=lfs merge=lfs -text
+src/ExpertLongBench-5-15-2025.png filter=lfs diff=lfs merge=lfs -text
+src/ExpertLongBench.png filter=lfs diff=lfs merge=lfs -text
+src/pipeline.png filter=lfs diff=lfs merge=lfs -text
+src/logo.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -1,13 +0,0 @@
-auto_evals/
-venv/
-__pycache__/
-.env
-.ipynb_checkpoints
-*ipynb
-.vscode/
-
-eval-queue/
-eval-results/
-eval-queue-bk/
-eval-results-bk/
-logs/

.pre-commit-config.yaml

@@ -1,53 +0,0 @@
-# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-default_language_version:
-  python: python3
-
-ci:
-  autofix_prs: true
-  autoupdate_commit_msg: '[pre-commit.ci] pre-commit suggestions'
-  autoupdate_schedule: quarterly
-
-repos:
-  - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.3.0
-    hooks:
-      - id: check-yaml
-      - id: check-case-conflict
-      - id: detect-private-key
-      - id: check-added-large-files
-        args: ['--maxkb=1000']
-      - id: requirements-txt-fixer
-      - id: end-of-file-fixer
-      - id: trailing-whitespace
-
-  - repo: https://github.com/PyCQA/isort
-    rev: 5.12.0
-    hooks:
-      - id: isort
-        name: Format imports
-
-  - repo: https://github.com/psf/black
-    rev: 22.12.0
-    hooks:
-      - id: black
-        name: Format code
-        additional_dependencies: ['click==8.0.2']
-
-  - repo: https://github.com/charliermarsh/ruff-pre-commit
-    # Ruff version.
-    rev: 'v0.0.267'
-    hooks:
-      - id: ruff

Dockerfile

@@ -0,0 +1,21 @@
+FROM python:3.9-slim
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    curl \
+    software-properties-common \
+    git \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt ./
+COPY src/ ./src/
+
+RUN pip3 install -r requirements.txt
+
+EXPOSE 8501
+
+HEALTHCHECK CMD curl --fail http://localhost:8501/_stcore/health
+
+ENTRYPOINT ["streamlit", "run", "src/streamlit_app.py", "--server.port=8501", "--server.address=0.0.0.0"]

Makefile

@@ -1,13 +0,0 @@
-.PHONY: style format
-
-
-style:
-	python -m black --line-length 119 .
-	python -m isort .
-	ruff check --fix .
-
-
-quality:
-	python -m black --check --line-length 119 .
-	python -m isort --check-only .
-	ruff check .

README.md

@@ -1,46 +1,20 @@
 ---
-title: ManyICLBench Leaderboard
-emoji: 🥇
-colorFrom: green
-colorTo: indigo
-sdk: gradio
-app_file: app.py
+title: ManyICLBench
+emoji: 🚀
+colorFrom: red
+colorTo: red
+sdk: docker
+app_port: 8501
+tags:
+- streamlit
 pinned: false
-license: mit
-short_description: Duplicate this leaderboard to initialize your own!
-sdk_version: 5.19.0
+short_description: Leaderboard for ManyICLBench
+license: cc-by-nc-4.0
 ---
 
-# Start the configuration
+# Welcome to Streamlit!
 
-Most of the variables to change for a default leaderboard are in `src/env.py` (replace the path for your leaderboard) and `src/about.py` (for tasks).
+Edit `/src/streamlit_app.py` to customize this app to your heart's desire. :heart:
 
-Results files should have the following format and be stored as json files:
-```json
-{
-    "config": {
-        "model_dtype": "torch.float16", # or torch.bfloat16 or 8bit or 4bit
-        "model_name": "path of the model on the hub: org/model",
-        "model_sha": "revision on the hub",
-    },
-    "results": {
-        "task_name": {
-            "metric_name": score,
-        },
-        "task_name2": {
-            "metric_name": score,
-        }
-    }
-}
-```
-
-Request files are created automatically by this tool.
-
-If you encounter problem on the space, don't hesitate to restart it to remove the create eval-queue, eval-queue-bk, eval-results and eval-results-bk created folder.
-
-# Code logic for more complex edits
-
-You'll find 
-- the main table' columns names and properties in `src/display/utils.py`
-- the logic to read all results and request files, then convert them in dataframe lines, in `src/leaderboard/read_evals.py`, and `src/populate.py`
-- the logic to allow or filter submissions in `src/submission/submit.py` and `src/submission/check_validity.py`
+If you have any questions, checkout our [documentation](https://docs.streamlit.io) and [community
+forums](https://discuss.streamlit.io).

app.py

@@ -1,204 +0,0 @@
-import gradio as gr
-from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns
-import pandas as pd
-from apscheduler.schedulers.background import BackgroundScheduler
-from huggingface_hub import snapshot_download
-
-from src.about import (
-    CITATION_BUTTON_LABEL,
-    CITATION_BUTTON_TEXT,
-    EVALUATION_QUEUE_TEXT,
-    INTRODUCTION_TEXT,
-    LLM_BENCHMARKS_TEXT,
-    TITLE,
-)
-from src.display.css_html_js import custom_css
-from src.display.utils import (
-    BENCHMARK_COLS,
-    COLS,
-    EVAL_COLS,
-    EVAL_TYPES,
-    AutoEvalColumn,
-    ModelType,
-    fields,
-    WeightType,
-    Precision
-)
-from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN
-from src.populate import get_evaluation_queue_df, get_leaderboard_df
-from src.submission.submit import add_new_eval
-
-
-def restart_space():
-    API.restart_space(repo_id=REPO_ID)
-
-### Space initialisation
-try:
-    print(EVAL_REQUESTS_PATH)
-    snapshot_download(
-        repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
-    )
-except Exception:
-    restart_space()
-try:
-    print(EVAL_RESULTS_PATH)
-    snapshot_download(
-        repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
-    )
-except Exception:
-    restart_space()
-
-
-LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
-
-(
-    finished_eval_queue_df,
-    running_eval_queue_df,
-    pending_eval_queue_df,
-) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
-
-def init_leaderboard(dataframe):
-    if dataframe is None or dataframe.empty:
-        raise ValueError("Leaderboard DataFrame is empty or None.")
-    return Leaderboard(
-        value=dataframe,
-        datatype=[c.type for c in fields(AutoEvalColumn)],
-        select_columns=SelectColumns(
-            default_selection=[c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],
-            cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden],
-            label="Select Columns to Display:",
-        ),
-        search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
-        hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden],
-        filter_columns=[
-            ColumnFilter(AutoEvalColumn.model_type.name, type="checkboxgroup", label="Model types"),
-            ColumnFilter(AutoEvalColumn.precision.name, type="checkboxgroup", label="Precision"),
-            ColumnFilter(
-                AutoEvalColumn.params.name,
-                type="slider",
-                min=0.01,
-                max=150,
-                label="Select the number of parameters (B)",
-            ),
-            ColumnFilter(
-                AutoEvalColumn.still_on_hub.name, type="boolean", label="Deleted/incomplete", default=True
-            ),
-        ],
-        bool_checkboxgroup_label="Hide models",
-        interactive=False,
-    )
-
-
-demo = gr.Blocks(css=custom_css)
-with demo:
-    gr.HTML(TITLE)
-    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
-
-    with gr.Tabs(elem_classes="tab-buttons") as tabs:
-        with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
-            leaderboard = init_leaderboard(LEADERBOARD_DF)
-
-        with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
-            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
-
-        with gr.TabItem("🚀 Submit here! ", elem_id="llm-benchmark-tab-table", id=3):
-            with gr.Column():
-                with gr.Row():
-                    gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")
-
-                with gr.Column():
-                    with gr.Accordion(
-                        f"✅ Finished Evaluations ({len(finished_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            finished_eval_table = gr.components.Dataframe(
-                                value=finished_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-                    with gr.Accordion(
-                        f"🔄 Running Evaluation Queue ({len(running_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            running_eval_table = gr.components.Dataframe(
-                                value=running_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-
-                    with gr.Accordion(
-                        f"⏳ Pending Evaluation Queue ({len(pending_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            pending_eval_table = gr.components.Dataframe(
-                                value=pending_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-            with gr.Row():
-                gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text")
-
-            with gr.Row():
-                with gr.Column():
-                    model_name_textbox = gr.Textbox(label="Model name")
-                    revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")
-                    model_type = gr.Dropdown(
-                        choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown],
-                        label="Model type",
-                        multiselect=False,
-                        value=None,
-                        interactive=True,
-                    )
-
-                with gr.Column():
-                    precision = gr.Dropdown(
-                        choices=[i.value.name for i in Precision if i != Precision.Unknown],
-                        label="Precision",
-                        multiselect=False,
-                        value="float16",
-                        interactive=True,
-                    )
-                    weight_type = gr.Dropdown(
-                        choices=[i.value.name for i in WeightType],
-                        label="Weights type",
-                        multiselect=False,
-                        value="Original",
-                        interactive=True,
-                    )
-                    base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)")
-
-            submit_button = gr.Button("Submit Eval")
-            submission_result = gr.Markdown()
-            submit_button.click(
-                add_new_eval,
-                [
-                    model_name_textbox,
-                    base_model_name_textbox,
-                    revision_name_textbox,
-                    precision,
-                    weight_type,
-                    model_type,
-                ],
-                submission_result,
-            )
-
-    with gr.Row():
-        with gr.Accordion("📙 Citation", open=False):
-            citation_button = gr.Textbox(
-                value=CITATION_BUTTON_TEXT,
-                label=CITATION_BUTTON_LABEL,
-                lines=20,
-                elem_id="citation-button",
-                show_copy_button=True,
-            )
-
-scheduler = BackgroundScheduler()
-scheduler.add_job(restart_space, "interval", seconds=1800)
-scheduler.start()
-demo.queue(default_concurrency_limit=40).launch()

pyproject.toml

@@ -1,13 +0,0 @@
-[tool.ruff]
-# Enable pycodestyle (`E`) and Pyflakes (`F`) codes by default.
-select = ["E", "F"]
-ignore = ["E501"] # line too long (black is taking care of this)
-line-length = 119
-fixable = ["A", "B", "C", "D", "E", "F", "G", "I", "N", "Q", "S", "T", "W", "ANN", "ARG", "BLE", "COM", "DJ", "DTZ", "EM", "ERA", "EXE", "FBT", "ICN", "INP", "ISC", "NPY", "PD", "PGH", "PIE", "PL", "PT", "PTH", "PYI", "RET", "RSE", "RUF", "SIM", "SLF", "TCH", "TID", "TRY", "UP", "YTT"]
-
-[tool.isort]
-profile = "black"
-line_length = 119
-
-[tool.black]
-line-length = 119

requirements.txt

@@ -1,16 +1,3 @@
-APScheduler
-black
-datasets
-gradio
-gradio[oauth]
-gradio_leaderboard==0.0.13
-gradio_client
-huggingface-hub>=0.18.0
-matplotlib
-numpy
+altair
 pandas
-python-dateutil
-tqdm
-transformers
-tokenizers>=0.15.0
-sentencepiece
+streamlit

src/Global Context Understanding_full_200.csv

@@ -0,0 +1,13 @@
+Task,1000,2000,4000,8000,16000,32000,64000,128000,avg,avg.L
+glm-4-9b-chat,40.50698427968745,40.282514030593525,42.03514758697741,42.77812732118466,40.698943674486365,40.46429917509213,38.84619005124577,39.12544210472614,40.59220602799918,39.47864377702135
+Mistral-Nemo-Instruct,38.25174117482217,39.06953517749783,39.28141110303884,38.98529668370538,33.061083213923375,32.830517013377566,30.45934094500344,27.110396670750852,34.88116524776493,30.133418209710623
+Mistral-Large-Instruct-AWQ,61.46838941271965,61.09587095162697,61.22583603295307,60.86833794699957,60.861933554427466,58.84010374503828,50.01197314627164,16.691663977868544,53.88301359598815,41.84791362305949
+Llama-3.1-8B-Instruct,37.30836868392976,38.84125527848912,41.2505736280164,40.794258067687366,39.83444694842053,39.773843356283,39.11869701930997,34.41209536535099,38.916692293435894,37.76821191364799
+Llama-3.1-70B-Instruct-AWQ,53.32457767581259,54.83670752732655,55.75854084271325,55.86647085146588,56.41586374288468,56.344722041867136,54.42144605517714,18.728575329808265,50.712113008381934,43.164914475617515
+Qwen2-7B-Instruct,39.521038462895056,41.95617309618548,45.16961923996148,45.388604614144,45.49983279321517,37.29314690113598,36.97410652200931,33.99279484585449,40.72441455942512,36.08668275633326
+Qwen2-72B-Instruct-AWQ,50.70543177597992,51.90185989326735,53.231917024118985,53.42772383454989,53.60456776354168,50.86853335158976,50.62023898985967,52.0547180736628,52.05187383832126,51.181163471704075
+Phi-3-Mini-Instruct,33.53774232920733,32.97066354630036,29.80053724275289,29.74568289344541,30.124298230078637,28.776180888367108,28.06044761555888,25.761550190503566,29.847137867026774,27.532726231476516
+Phi-3-Medium-Instruct,41.587771292424094,40.912198239966216,34.854733073016845,35.634932868512294,36.90794452772758,36.83888019383616,36.37856781039721,28.30918854780435,36.42802706921059,33.842212184012574
+Phi-3-Small-Instruct,41.61109509590759,41.61237571572385,41.606181105263296,35.57902498126983,37.16841325462178,37.72671633709681,36.90648845260399,35.328537592998906,38.442354066935756,36.653914127566566
+Jamba-1.5-Mini,31.96393502346349,33.0804151894422,32.97355597731929,32.70266682086676,31.65788220876118,28.82458742147524,27.14127682848188,25.874333552669743,30.527331627809975,27.280065934208952
+Gemini-1.5-Pro,57.87,63.39,64.15,66.78,68.02,67.78,66.14,66.42,65.07,66.78

src/Retrieval_full_200.csv

@@ -0,0 +1,13 @@
+Task,1000,2000,4000,8000,16000,32000,64000,128000,avg,avg.L
+glm-4-9b-chat,31.634064359831303,34.989931742075846,46.36923333133574,57.26961338964374,63.61295034310319,68.33923257857671,72.15825322852012,72.92730988128146,55.91257360679601,71.14159856279277
+Mistral-Nemo-Instruct,33.442943349483386,35.45098716112904,48.17141258583688,57.95087475137824,65.37779827704927,65.49015807503481,63.60618294804688,61.730166764004046,53.90256548899532,63.60883592902858
+Mistral-Large-Instruct-AWQ,49.14942052867251,51.23221381539887,60.7793522691761,71.94684170889438,77.09652619929832,79.45080285304805,77.76925084338332,62.18320059627511,66.20095110176834,73.13441809756883
+Llama-3.1-8B-Instruct,32.40198895196592,32.979539690235654,45.0106861254419,56.07951990889208,64.18069403777689,69.62343477623669,71.53647862646864,69.22900618587121,55.130168537861124,70.12963986285885
+Llama-3.1-70B-Instruct-AWQ,38.7480657800462,42.87159365068116,53.97630601173803,66.0747722354859,73.11797388808664,76.55703704352354,78.2577157921391,65.5267461115416,61.89127631415526,73.44716631573475
+Qwen2-7B-Instruct,31.530664252548508,35.5910761564275,44.63338088737434,54.35904179697225,63.05187841718273,64.46105522821793,66.9059768609517,63.06871149116235,52.95022313635466,64.81191452677733
+Qwen2-72B-Instruct-AWQ,36.40772920298174,41.89206817510168,54.241398436281806,65.33266907435282,73.39122500043426,76.52590896239253,77.50975041531201,77.46999427458152,62.8463429426798,77.16855121742869
+Phi-3-Mini-Instruct,30.26814474962816,30.90238348500332,38.09444166155989,48.140298057357256,53.58382886518213,57.28513647576864,56.83349679389524,48.723036083017135,45.47884577142647,54.28055645089368
+Phi-3-Medium-Instruct,31.72976117411381,33.54934540740912,39.10262687186865,49.82711345124473,58.29010994509822,61.17171452761886,60.62625650790335,45.31693119039325,47.45173238445625,55.70496740863848
+Phi-3-Small-Instruct,31.48315960873518,36.27405424169346,46.19687960118542,54.34055832812775,59.62975835311172,59.733406453060674,60.19785010451274,48.96685678550229,49.60281543449115,56.299371114358564
+Jamba-1.5-Mini,32.100142207922914,36.90890463687818,48.61257141384665,60.28550585861812,66.0507948010125,68.33428275575879,66.01898750117914,65.27696038728068,55.448518695312124,66.54341021473954
+Gemini-1.5-Pro,36.40,47.31,58.01,65.49,71.43,74.22,72.43,72.42,62.21,73.03

src/about.py

@@ -1,72 +0,0 @@
-from dataclasses import dataclass
-from enum import Enum
-
-@dataclass
-class Task:
-    benchmark: str
-    metric: str
-    col_name: str
-
-
-# Select your tasks here
-# ---------------------------------------------------
-class Tasks(Enum):
-    # task_key in the json file, metric_key in the json file, name to display in the leaderboard 
-    task0 = Task("anli_r1", "acc", "ANLI")
-    task1 = Task("logiqa", "acc_norm", "LogiQA")
-
-NUM_FEWSHOT = 0 # Change with your few shot
-# ---------------------------------------------------
-
-
-
-# Your leaderboard name
-TITLE = """<h1 align="center" id="space-title">Demo leaderboard</h1>"""
-
-# What does your leaderboard evaluate?
-INTRODUCTION_TEXT = """
-Intro text
-"""
-
-# Which evaluations are you running? how can people reproduce what you have?
-LLM_BENCHMARKS_TEXT = f"""
-## How it works
-
-## Reproducibility
-To reproduce our results, here is the commands you can run:
-
-"""
-
-EVALUATION_QUEUE_TEXT = """
-## Some good practices before submitting a model
-
-### 1) Make sure you can load your model and tokenizer using AutoClasses:
-```python
-from transformers import AutoConfig, AutoModel, AutoTokenizer
-config = AutoConfig.from_pretrained("your model name", revision=revision)
-model = AutoModel.from_pretrained("your model name", revision=revision)
-tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
-```
-If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded.
-
-Note: make sure your model is public!
-Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!
-
-### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
-It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!
-
-### 3) Make sure your model has an open license!
-This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗
-
-### 4) Fill up your model card
-When we add extra information about models to the leaderboard, it will be automatically taken from the model card
-
-## In case of model failure
-If your model is displayed in the `FAILED` category, its execution stopped.
-Make sure you have followed the above steps first.
-If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task).
-"""
-
-CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
-CITATION_BUTTON_TEXT = r"""
-"""

src/display/css_html_js.py

@@ -1,105 +0,0 @@
-custom_css = """
-
-.markdown-text {
-    font-size: 16px !important;
-}
-
-#models-to-add-text {
-    font-size: 18px !important;
-}
-
-#citation-button span {
-    font-size: 16px !important;
-}
-
-#citation-button textarea {
-    font-size: 16px !important;
-}
-
-#citation-button > label > button {
-    margin: 6px;
-    transform: scale(1.3);
-}
-
-#leaderboard-table {
-    margin-top: 15px
-}
-
-#leaderboard-table-lite {
-    margin-top: 15px
-}
-
-#search-bar-table-box > div:first-child {
-    background: none;
-    border: none;
-}
- 
-#search-bar {
-    padding: 0px;
-}
-
-/* Limit the width of the first AutoEvalColumn so that names don't expand too much */
-#leaderboard-table td:nth-child(2),
-#leaderboard-table th:nth-child(2) {
-    max-width: 400px;
-    overflow: auto;
-    white-space: nowrap;
-}
-
-.tab-buttons button {
-    font-size: 20px;
-}
-
-#scale-logo {
-    border-style: none !important;
-    box-shadow: none;
-    display: block;
-    margin-left: auto;
-    margin-right: auto;
-    max-width: 600px;
-}
-
-#scale-logo .download {
-    display: none;
-}
-#filter_type{
-    border: 0;
-    padding-left: 0;
-    padding-top: 0;
-}
-#filter_type label {
-    display: flex;
-}
-#filter_type label > span{
-    margin-top: var(--spacing-lg);
-    margin-right: 0.5em;
-}
-#filter_type label > .wrap{
-    width: 103px;
-}
-#filter_type label > .wrap .wrap-inner{  
-    padding: 2px;
-}
-#filter_type label > .wrap .wrap-inner input{
-    width: 1px
-}
-#filter-columns-type{
-    border:0;
-    padding:0.5;
-}
-#filter-columns-size{
-    border:0;
-    padding:0.5;
-}
-#box-filter > .form{
-    border: 0
-}
-"""
-
-get_window_url_params = """
-    function(url_params) {
-        const params = new URLSearchParams(window.location.search);
-        url_params = Object.fromEntries(params);
-        return url_params;
-    }
-    """

src/display/formatting.py

@@ -1,27 +0,0 @@
-def model_hyperlink(link, model_name):
-    return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
-
-
-def make_clickable_model(model_name):
-    link = f"https://huggingface.co/{model_name}"
-    return model_hyperlink(link, model_name)
-
-
-def styled_error(error):
-    return f"<p style='color: red; font-size: 20px; text-align: center;'>{error}</p>"
-
-
-def styled_warning(warn):
-    return f"<p style='color: orange; font-size: 20px; text-align: center;'>{warn}</p>"
-
-
-def styled_message(message):
-    return f"<p style='color: green; font-size: 20px; text-align: center;'>{message}</p>"
-
-
-def has_no_nan_values(df, columns):
-    return df[columns].notna().all(axis=1)
-
-
-def has_nan_values(df, columns):
-    return df[columns].isna().any(axis=1)

src/display/utils.py

@@ -1,110 +0,0 @@
-from dataclasses import dataclass, make_dataclass
-from enum import Enum
-
-import pandas as pd
-
-from src.about import Tasks
-
-def fields(raw_class):
-    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]
-
-
-# These classes are for user facing column names,
-# to avoid having to change them all around the code
-# when a modif is needed
-@dataclass
-class ColumnContent:
-    name: str
-    type: str
-    displayed_by_default: bool
-    hidden: bool = False
-    never_hidden: bool = False
-
-## Leaderboard columns
-auto_eval_column_dict = []
-# Init
-auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
-auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
-#Scores
-auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])
-for task in Tasks:
-    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
-# Model information
-auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
-auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
-auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
-auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
-auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])
-auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
-auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])
-auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False)])
-auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
-
-# We use make dataclass to dynamically fill the scores from Tasks
-AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)
-
-## For the queue columns in the submission tab
-@dataclass(frozen=True)
-class EvalQueueColumn:  # Queue column
-    model = ColumnContent("model", "markdown", True)
-    revision = ColumnContent("revision", "str", True)
-    private = ColumnContent("private", "bool", True)
-    precision = ColumnContent("precision", "str", True)
-    weight_type = ColumnContent("weight_type", "str", "Original")
-    status = ColumnContent("status", "str", True)
-
-## All the model information that we might need
-@dataclass
-class ModelDetails:
-    name: str
-    display_name: str = ""
-    symbol: str = "" # emoji
-
-
-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="🟦")
-    Unknown = ModelDetails(name="", symbol="?")
-
-    def to_str(self, separator=" "):
-        return f"{self.value.symbol}{separator}{self.value.name}"
-
-    @staticmethod
-    def from_str(type):
-        if "fine-tuned" in type or "🔶" in type:
-            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
-        return ModelType.Unknown
-
-class WeightType(Enum):
-    Adapter = ModelDetails("Adapter")
-    Original = ModelDetails("Original")
-    Delta = ModelDetails("Delta")
-
-class Precision(Enum):
-    float16 = ModelDetails("float16")
-    bfloat16 = ModelDetails("bfloat16")
-    Unknown = ModelDetails("?")
-
-    def from_str(precision):
-        if precision in ["torch.float16", "float16"]:
-            return Precision.float16
-        if precision in ["torch.bfloat16", "bfloat16"]:
-            return Precision.bfloat16
-        return Precision.Unknown
-
-# Column selection
-COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
-
-EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
-EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]
-
-BENCHMARK_COLS = [t.value.col_name for t in Tasks]
-

src/envs.py

@@ -1,25 +0,0 @@
-import os
-
-from huggingface_hub import HfApi
-
-# Info to change for your repository
-# ----------------------------------
-TOKEN = os.environ.get("HF_TOKEN") # A read/write token for your org
-
-OWNER = "demo-leaderboard-backend" # Change to your org - don't forget to create a results and request dataset, with the correct format!
-# ----------------------------------
-
-REPO_ID = f"{OWNER}/leaderboard"
-QUEUE_REPO = f"{OWNER}/requests"
-RESULTS_REPO = f"{OWNER}/results"
-
-# If you setup a cache later, just change HF_HOME
-CACHE_PATH=os.getenv("HF_HOME", ".")
-
-# Local caches
-EVAL_REQUESTS_PATH = os.path.join(CACHE_PATH, "eval-queue")
-EVAL_RESULTS_PATH = os.path.join(CACHE_PATH, "eval-results")
-EVAL_REQUESTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-queue-bk")
-EVAL_RESULTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-results-bk")
-
-API = HfApi(token=TOKEN)

src/leaderboard/read_evals.py

@@ -1,196 +0,0 @@
-import glob
-import json
-import math
-import os
-from dataclasses import dataclass
-
-import dateutil
-import numpy as np
-
-from src.display.formatting import make_clickable_model
-from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType
-from src.submission.check_validity import is_model_on_hub
-
-
-@dataclass
-class EvalResult:
-    """Represents one full evaluation. Built from a combination of the result and request file for a given run.
-    """
-    eval_name: str # org_model_precision (uid)
-    full_model: str # org/model (path on hub)
-    org: str 
-    model: str
-    revision: str # commit hash, "" if main
-    results: dict
-    precision: Precision = Precision.Unknown
-    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
-    weight_type: WeightType = WeightType.Original # Original or Adapter
-    architecture: str = "Unknown" 
-    license: str = "?"
-    likes: int = 0
-    num_params: int = 0
-    date: str = "" # submission date of request file
-    still_on_hub: bool = False
-
-    @classmethod
-    def init_from_json_file(self, json_filepath):
-        """Inits the result from the specific model result file"""
-        with open(json_filepath) as fp:
-            data = json.load(fp)
-
-        config = data.get("config")
-
-        # Precision
-        precision = Precision.from_str(config.get("model_dtype"))
-
-        # Get model and org
-        org_and_model = config.get("model_name", config.get("model_args", None))
-        org_and_model = org_and_model.split("/", 1)
-
-        if len(org_and_model) == 1:
-            org = None
-            model = org_and_model[0]
-            result_key = f"{model}_{precision.value.name}"
-        else:
-            org = org_and_model[0]
-            model = org_and_model[1]
-            result_key = f"{org}_{model}_{precision.value.name}"
-        full_model = "/".join(org_and_model)
-
-        still_on_hub, _, model_config = is_model_on_hub(
-            full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False
-        )
-        architecture = "?"
-        if model_config is not None:
-            architectures = getattr(model_config, "architectures", None)
-            if architectures:
-                architecture = ";".join(architectures)
-
-        # Extract results available in this file (some results are split in several files)
-        results = {}
-        for task in Tasks:
-            task = task.value
-
-            # We average all scores of a given metric (not all metrics are present in all files)
-            accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k])
-            if accs.size == 0 or any([acc is None for acc in accs]):
-                continue
-
-            mean_acc = np.mean(accs) * 100.0
-            results[task.benchmark] = mean_acc
-
-        return self(
-            eval_name=result_key,
-            full_model=full_model,
-            org=org,
-            model=model,
-            results=results,
-            precision=precision,  
-            revision= config.get("model_sha", ""),
-            still_on_hub=still_on_hub,
-            architecture=architecture
-        )
-
-    def update_with_request_file(self, requests_path):
-        """Finds the relevant request file for the current model and updates info with it"""
-        request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)
-
-        try:
-            with open(request_file, "r") as f:
-                request = json.load(f)
-            self.model_type = ModelType.from_str(request.get("model_type", ""))
-            self.weight_type = WeightType[request.get("weight_type", "Original")]
-            self.license = request.get("license", "?")
-            self.likes = request.get("likes", 0)
-            self.num_params = request.get("params", 0)
-            self.date = request.get("submitted_time", "")
-        except Exception:
-            print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")
-
-    def to_dict(self):
-        """Converts the Eval Result to a dict compatible with our dataframe display"""
-        average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
-        data_dict = {
-            "eval_name": self.eval_name,  # not a column, just a save name,
-            AutoEvalColumn.precision.name: self.precision.value.name,
-            AutoEvalColumn.model_type.name: self.model_type.value.name,
-            AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
-            AutoEvalColumn.weight_type.name: self.weight_type.value.name,
-            AutoEvalColumn.architecture.name: self.architecture,
-            AutoEvalColumn.model.name: make_clickable_model(self.full_model),
-            AutoEvalColumn.revision.name: self.revision,
-            AutoEvalColumn.average.name: average,
-            AutoEvalColumn.license.name: self.license,
-            AutoEvalColumn.likes.name: self.likes,
-            AutoEvalColumn.params.name: self.num_params,
-            AutoEvalColumn.still_on_hub.name: self.still_on_hub,
-        }
-
-        for task in Tasks:
-            data_dict[task.value.col_name] = self.results[task.value.benchmark]
-
-        return data_dict
-
-
-def get_request_file_for_model(requests_path, model_name, precision):
-    """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
-    request_files = os.path.join(
-        requests_path,
-        f"{model_name}_eval_request_*.json",
-    )
-    request_files = glob.glob(request_files)
-
-    # Select correct request file (precision)
-    request_file = ""
-    request_files = sorted(request_files, reverse=True)
-    for tmp_request_file in request_files:
-        with open(tmp_request_file, "r") as f:
-            req_content = json.load(f)
-            if (
-                req_content["status"] in ["FINISHED"]
-                and req_content["precision"] == precision.split(".")[-1]
-            ):
-                request_file = tmp_request_file
-    return request_file
-
-
-def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:
-    """From the path of the results folder root, extract all needed info for results"""
-    model_result_filepaths = []
-
-    for root, _, files in os.walk(results_path):
-        # We should only have json files in model results
-        if len(files) == 0 or any([not f.endswith(".json") for f in files]):
-            continue
-
-        # Sort the files by date
-        try:
-            files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
-        except dateutil.parser._parser.ParserError:
-            files = [files[-1]]
-
-        for file in files:
-            model_result_filepaths.append(os.path.join(root, file))
-
-    eval_results = {}
-    for model_result_filepath in model_result_filepaths:
-        # Creation of result
-        eval_result = EvalResult.init_from_json_file(model_result_filepath)
-        eval_result.update_with_request_file(requests_path)
-
-        # Store results of same eval together
-        eval_name = eval_result.eval_name
-        if eval_name in eval_results.keys():
-            eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
-        else:
-            eval_results[eval_name] = eval_result
-
-    results = []
-    for v in eval_results.values():
-        try:
-            v.to_dict() # we test if the dict version is complete
-            results.append(v)
-        except KeyError:  # not all eval values present
-            continue
-
-    return results

src/populate.py

@@ -1,58 +0,0 @@
-import json
-import os
-
-import pandas as pd
-
-from src.display.formatting import has_no_nan_values, make_clickable_model
-from src.display.utils import AutoEvalColumn, EvalQueueColumn
-from src.leaderboard.read_evals import get_raw_eval_results
-
-
-def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
-    """Creates a dataframe from all the individual experiment results"""
-    raw_data = get_raw_eval_results(results_path, requests_path)
-    all_data_json = [v.to_dict() for v in raw_data]
-
-    df = pd.DataFrame.from_records(all_data_json)
-    df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
-    df = df[cols].round(decimals=2)
-
-    # filter out if any of the benchmarks have not been produced
-    df = df[has_no_nan_values(df, benchmark_cols)]
-    return df
-
-
-def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
-    """Creates the different dataframes for the evaluation queues requestes"""
-    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
-    all_evals = []
-
-    for entry in entries:
-        if ".json" in entry:
-            file_path = os.path.join(save_path, entry)
-            with open(file_path) as fp:
-                data = json.load(fp)
-
-            data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
-            data[EvalQueueColumn.revision.name] = data.get("revision", "main")
-
-            all_evals.append(data)
-        elif ".md" not in entry:
-            # this is a folder
-            sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")]
-            for sub_entry in sub_entries:
-                file_path = os.path.join(save_path, entry, sub_entry)
-                with open(file_path) as fp:
-                    data = json.load(fp)
-
-                data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
-                data[EvalQueueColumn.revision.name] = data.get("revision", "main")
-                all_evals.append(data)
-
-    pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]
-    running_list = [e for e in all_evals if e["status"] == "RUNNING"]
-    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]
-    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
-    df_running = pd.DataFrame.from_records(running_list, columns=cols)
-    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
-    return df_finished[cols], df_running[cols], df_pending[cols]

src/streamlit_app.py

@@ -0,0 +1,222 @@
+import streamlit as st
+import pandas as pd
+from PIL import Image
+import base64
+from io import BytesIO  
+import numpy as np
+
+# ─── Page config ──────────────────────────────────────────────────────────────
+st.set_page_config(page_title="ManyICLBench Leaderboard", layout="wide")
+logo_image = Image.open("src/manyicl_logo.png")
+
+
+def encode_image(image):
+    buffered = BytesIO()
+    image.save(buffered, format="PNG")
+    return base64.b64encode(buffered.getvalue()).decode("utf-8")
+
+img_data = encode_image(logo_image)
+
+st.markdown(
+    f"""
+    <div class="logo-container" style="display:flex; justify-content: center; align-items: center; gap: 20px;">
+        <img src="data:image/png;base64,{img_data}" style="width:50%; max-width:700px;"/>
+    </div>
+    """,
+    unsafe_allow_html=True
+)
+
+st.markdown(
+    '''
+    <div class="header">
+        <br/>
+        <p style="font-size:22px;">
+        ManyICLBench: Benchmarking Large Language Models' Long Context Capabilities with Many-Shot In-Context Learning
+        </p>
+        <p style="font-size:20px;">
+              📑 <a href="https://arxiv.org/abs/2411.07130">Paper</a> | 💻 <a href="https://github.com/launchnlp/ManyICLBench">GitHub</a> | 🤗 <a href="https://huggingface.co/datasets/launch/ManyICLBench/">Dataset</a> |     
+               ⚙️ <strong>Version</strong>: <strong>V1</strong> | <strong># Models</strong>: 12 | Updated: <strong>June 2025</strong>
+        </p>
+    </div>
+    ''',
+    unsafe_allow_html=True
+)
+
+# ─── Load data ────────────────────────────────────────────────────────────────
+@st.cache_data
+def load_data(path):
+    df = pd.read_csv(path)
+    if 'Task' in df.columns:  # Rename Task to Models for consistency
+        df = df.rename(columns={'Task': 'Models'})
+    score_cols = ['1000', '2000', '4000', '8000', '16000', '32000', '64000', '128000']
+    # Keep existing avg and avg.L columns
+    # Compute rank per column (1 = best)
+    for col in score_cols + ['avg', 'avg.L']:
+        df[f"{col}_rank"] = df[col].rank(ascending=False, method="min").astype(int)
+    return df
+
+# Add evaluation metrics explanation
+st.markdown("## 📊 Evaluation Metrics")
+st.markdown("""
+- **Per-length Performance**: Performance at different context lengths (1K to 128K tokens)
+- **avg**: Average performance across all context lengths
+- **avg.L**: Average performance on longer contexts (>32K tokens)
+
+Higher scores indicate better performance, with all metrics reported as percentages (0-100).
+
+Red indicates performance improvement compared to 1k. Blue indicates performance downgrade compared to 1k. A darker color means higher improvement or downgrade.
+""")
+
+def display_table(df, cols):
+    # Precompute max values for Avg and Avg.L
+    max_avg = df['avg'].max()
+    max_avg_l = df['avg.L'].max()
+
+    # Build raw HTML table
+    html = "<table style='border-collapse:collapse; width:100%; font-size:14px;'>"
+    
+    # Format header labels
+    html += "<tr>"
+    for col in cols:
+        style = "padding:6px;"
+        label = ""
+        if col in ['1000', '2000', '4000', '8000', '16000', '32000', '64000', '128000']:
+            # Convert to K format
+            val = int(col) // 1000
+            label = f"{val}K"
+        else:
+            label = col.title()  # Capitalize first letter
+        
+        if col in ["Model", "Models"]:
+            style += " width: 15%;"
+
+        html += f"<th style='{style}'>{label}</th>"
+    html += "</tr>"
+    
+    # rows
+    for _, row in df.iterrows():
+        html += "<tr>"
+        for col in cols:
+            val = row[col]
+            if col in ["Model", "Models"]:
+                html += f"<td style='padding:6px; text-align:left; width: 15%;'>{val}</td>"
+            else:
+                # Format value
+                val_str = f"{val:.1f}" if isinstance(val, (float, np.float64)) else val
+                
+                # Determine if this column should be colored
+                if col in ['1000', 'avg', 'avg.L']:
+                    # No coloring for these columns, but add bolding for max values
+                    bold = ""
+                    if (col == 'avg' and val == max_avg) or \
+                       (col == 'avg.L' and val == max_avg_l):
+                        bold = "font-weight:bold;"
+                    style = f"padding:6px; border: 1px solid #444; {bold}"
+                else:
+                    # Calculate relative improvement from 1k baseline
+                    baseline = float(row['1000'])
+                    if baseline != 0:
+                        relative_change = float(val) / baseline - 1  # -1 to center at 0
+                        # Clamp the change to a reasonable range for color scaling
+                        clamped_change = max(min(relative_change, 1.5), -0.5)
+                        
+                        # Normalize to 0-1 range where 0.5 is the neutral point (no change)
+                        if clamped_change < 0:
+                            # Map [-0.5, 0) to [0, 0.5)
+                            norm = clamped_change + 0.5
+                        else:
+                            # Map [0, 1.5] to [0.5, 1.0]
+                            norm = 0.5 + (clamped_change / 3.0)
+
+                        # Color interpolation:
+                        # norm = 0   -> blue  (100, 149, 237)
+                        # norm = 0.5 -> white (255, 255, 255)
+                        # norm = 1   -> red   (220, 20, 60)
+                        
+                        if norm < 0.5:  # Blue to White
+                            # Interpolate from blue to white
+                            factor = norm * 2  # 0 to 1
+                            r = int(100 + (255 - 100) * factor)
+                            g = int(149 + (255 - 149) * factor)
+                            b = int(237 + (255 - 237) * factor)
+                        else:  # White to Red
+                            # Interpolate from white to red
+                            factor = (norm - 0.5) * 2  # 0 to 1
+                            r = int(255 - (255 - 220) * factor)
+                            g = int(255 - (255 - 20) * factor)
+                            b = int(255 - (255 - 60) * factor)
+                        
+                        style = f"background-color:rgba({r},{g},{b},0.8); padding:6px; border: 1px solid #444;"
+                    else:
+                        style = "padding:6px; border: 1px solid #444;"
+                
+                html += f"<td style='{style}'>{val_str}</td>"
+        html += "</tr>"
+    html += "</table>"
+    st.markdown(html, unsafe_allow_html=True)
+
+# Display Retrieval table
+st.markdown("## SSL Tasks")
+st.markdown("Similar-sample Learning tasks require models to learn from a small set of similar demostration, therefore evaluating models' ability to retrieve similar samples.")
+df = load_data("src/Retrieval_full_200.csv")
+cols = ["Models", "1000", "2000", "4000", "8000", "16000", "32000", "64000", "128000", "avg", "avg.L"]
+display_table(df, cols)
+
+# Display Global Context Understanding table
+st.markdown("## ASL Tasks")
+st.markdown("All-Sample Learning tasks require models to learn from all the demostrations, therefore evaluating models' ability to understand the global context.")
+df = load_data("src/Global Context Understanding_full_200.csv")
+cols = ["Models", "1000", "2000", "4000", "8000", "16000", "32000", "64000", "128000", "avg", "avg.L"]
+display_table(df, cols)
+
+st.markdown("## 📚 Abstract")
+st.write(
+"""
+Many-shot in-context learning (ICL) has emerged as a unique setup to both utilize and test the ability of large language models to handle long context.This paper delves into long-context language model (LCLM) evaluation through many-shot ICL. We first ask: what types of ICL tasks benefit from additional demonstrations, and how effective are they in evaluating LCLMs?
+We find that classification and summarization tasks show performance improvements with additional demonstrations, while translation and reasoning tasks do not exhibit clear trends. 
+Next, we investigate the extent to which different tasks necessitate retrieval versus global context understanding.
+We develop metrics to categorize ICL tasks into two groups: (i) similar-sample learning (**SSL**): tasks where retrieval of the most similar examples is sufficient for good performance, and (ii) all-sample learning (**ASL**): tasks that necessitate a deeper comprehension of all examples in the prompt. 
+Lastly, we introduce a new many-shot ICL benchmark built on existing ICL tasks, ManyICLBench, to characterize model's ability on both fronts and benchmark 12 LCLMs using ManyICLBench. We find that while state-of-the-art models demonstrate good performance up to 64k tokens in SSL tasks, many models experience significant performance drops at only 16k tokens in ASL tasks.
+"""
+)
+st.markdown("## Dataset Details")
+st.markdown("""
+| **Dataset** | **Task Category** | **Avg. Tokens / Shot** | **Max # of Shots** | **# of Tasks** |
+| :--- | :--- | :--- | :--- | :--- |
+| BANKING77 | Intent Classification | 13.13 | 5386 | 1 |
+| GoEmotions | Emotion Classification | 15.85 | 5480 | 1 |
+| DialogRE | Relation Classification | 233.27 | 395 | 1 |
+| TREC | Question Classification | 11.25 | 6272 | 1 |
+| CLINC150 | Intent Classification | 8.95 | 7252 | 1 |
+| MATH | Math reasoning | [185.52, 407.90] | [286, 653] | 4 |
+| GSM8K | Math reasoning | 55.78 | 784 | 1 |
+| BBH | Reasoning | [48.27, 243.01] | [406, 2660] | 4 |
+| GPQA | MQ - Science | [183.55, 367.02] | [314, 580] | 1 |
+| ARC | MQ - Science | [61.54, 61.54] | [1997, 2301] | 2 |
+| XLSUM | New Summarization | 621.32 | 220 | 1 |
+
+GPT-4o tokenizer is used to calculate # of tokens. Max # of shots is the number of shots can be fitted into the 128k context window. For datasets that have multiple subtasks, we list the range for each value.
+
+**ASL Tasks**: banking77, dialogRE, TREC, CLINC150, and BBH_geometric_shapes
+
+**SSL Tasks**: GSM8K, MATH tasks, GSM8K, XLSUM, GPQA_cot, ARC_challenge, BBH-dyck_languages, BBH-salient_translation_error_detection, and BBH-word_sorting.
+
+""")
+st.markdown('## 🤖 Submit Your Model')
+st.write(
+"""
+👉 You can submit your model through the following link: [https://forms.gle/eWjzPDusDJSbXCCT7](https://forms.gle/eWjzPDusDJSbXCCT7)
+"""
+)
+
+st.markdown("## 📚 Citation")
+st.write("""
+```bibtex
+@article{zou2025manyshotincontextlearninglongcontext,
+  title={On Many-Shot In-Context Learning for Long-Context Evaluation}, 
+  author={Kaijian Zou and Muhammad Khalifa and Lu Wang},
+  journal={arXiv preprint arXiv:2411.07130},
+  year={2025}
+}
+```
+""")

src/submission/check_validity.py

@@ -1,99 +0,0 @@
-import json
-import os
-import re
-from collections import defaultdict
-from datetime import datetime, timedelta, timezone
-
-import huggingface_hub
-from huggingface_hub import ModelCard
-from huggingface_hub.hf_api import ModelInfo
-from transformers import AutoConfig
-from transformers.models.auto.tokenization_auto import AutoTokenizer
-
-def check_model_card(repo_id: str) -> tuple[bool, str]:
-    """Checks if the model card and license exist and have been filled"""
-    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."
-
-    # Enforce license metadata
-    if card.data.license is None:
-        if not ("license_name" in card.data and "license_link" in card.data):
-            return False, (
-                "License not found. Please add a license to your model card using the `license` metadata or a"
-                " `license_name`/`license_link` pair."
-            )
-
-    # Enforce card content
-    if len(card.text) < 200:
-        return False, "Please add a description to your model card, it is too short."
-
-    return True, ""
-
-def is_model_on_hub(model_name: str, revision: str, token: str = None, trust_remote_code=False, test_tokenizer=False) -> tuple[bool, str]:
-    """Checks if the model model_name is on the hub, and whether it (and its tokenizer) can be loaded with AutoClasses."""
-    try:
-        config = AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
-        if test_tokenizer:
-            try:
-                tk = AutoTokenizer.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
-            except ValueError as e:
-                return (
-                    False,
-                    f"uses a tokenizer which is not in a transformers release: {e}",
-                    None
-                )
-            except Exception as e:
-                return (False, "'s tokenizer cannot be loaded. Is your tokenizer class in a stable transformers release, and correctly configured?", None)
-        return True, None, config
-
-    except ValueError:
-        return (
-            False,
-            "needs to be launched with `trust_remote_code=True`. For safety reason, we do not allow these models to be automatically submitted to the leaderboard.",
-            None
-        )
-
-    except Exception as e:
-        return False, "was not found on hub!", None
-
-
-def get_model_size(model_info: ModelInfo, precision: str):
-    """Gets the model size from the configuration, or the model name if the configuration does not contain the information."""
-    try:
-        model_size = round(model_info.safetensors["total"] / 1e9, 3)
-    except (AttributeError, TypeError):
-        return 0  # Unknown model sizes are indicated as 0, see NUMERIC_INTERVALS in app.py
-
-    size_factor = 8 if (precision == "GPTQ" or "gptq" in model_info.modelId.lower()) else 1
-    model_size = size_factor * model_size
-    return model_size
-
-def get_model_arch(model_info: ModelInfo):
-    """Gets the model architecture from the configuration"""
-    return model_info.config.get("architectures", "Unknown")
-
-def already_submitted_models(requested_models_dir: str) -> set[str]:
-    """Gather a list of already submitted models to avoid duplicates"""
-    depth = 1
-    file_names = []
-    users_to_submission_dates = defaultdict(list)
-
-    for root, _, files in os.walk(requested_models_dir):
-        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
-        if current_depth == depth:
-            for file in files:
-                if not file.endswith(".json"):
-                    continue
-                with open(os.path.join(root, file), "r") as f:
-                    info = json.load(f)
-                    file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}")
-
-                    # Select organisation
-                    if info["model"].count("/") == 0 or "submitted_time" not in info:
-                        continue
-                    organisation, _ = info["model"].split("/")
-                    users_to_submission_dates[organisation].append(info["submitted_time"])
-
-    return set(file_names), users_to_submission_dates

src/submission/submit.py

@@ -1,119 +0,0 @@
-import json
-import os
-from datetime import datetime, timezone
-
-from src.display.formatting import styled_error, styled_message, styled_warning
-from src.envs import API, EVAL_REQUESTS_PATH, TOKEN, QUEUE_REPO
-from src.submission.check_validity import (
-    already_submitted_models,
-    check_model_card,
-    get_model_size,
-    is_model_on_hub,
-)
-
-REQUESTED_MODELS = None
-USERS_TO_SUBMISSION_DATES = None
-
-def add_new_eval(
-    model: str,
-    base_model: str,
-    revision: str,
-    precision: str,
-    weight_type: str,
-    model_type: str,
-):
-    global REQUESTED_MODELS
-    global USERS_TO_SUBMISSION_DATES
-    if not REQUESTED_MODELS:
-        REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = already_submitted_models(EVAL_REQUESTS_PATH)
-
-    user_name = ""
-    model_path = model
-    if "/" in model:
-        user_name = model.split("/")[0]
-        model_path = model.split("/")[1]
-
-    precision = precision.split(" ")[0]
-    current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
-
-    if model_type is None or model_type == "":
-        return styled_error("Please select a model type.")
-
-    # Does the model actually exist?
-    if revision == "":
-        revision = "main"
-
-    # Is the model on the hub?
-    if weight_type in ["Delta", "Adapter"]:
-        base_model_on_hub, error, _ = is_model_on_hub(model_name=base_model, revision=revision, token=TOKEN, test_tokenizer=True)
-        if not base_model_on_hub:
-            return styled_error(f'Base model "{base_model}" {error}')
-
-    if not weight_type == "Adapter":
-        model_on_hub, error, _ = is_model_on_hub(model_name=model, revision=revision, token=TOKEN, test_tokenizer=True)
-        if not model_on_hub:
-            return styled_error(f'Model "{model}" {error}')
-
-    # Is the model info correctly filled?
-    try:
-        model_info = API.model_info(repo_id=model, revision=revision)
-    except Exception:
-        return styled_error("Could not get your model information. Please fill it up properly.")
-
-    model_size = get_model_size(model_info=model_info, precision=precision)
-
-    # Were the model card and license filled?
-    try:
-        license = model_info.cardData["license"]
-    except Exception:
-        return styled_error("Please select a license for your model")
-
-    modelcard_OK, error_msg = check_model_card(model)
-    if not modelcard_OK:
-        return styled_error(error_msg)
-
-    # Seems good, creating the eval
-    print("Adding new eval")
-
-    eval_entry = {
-        "model": model,
-        "base_model": base_model,
-        "revision": revision,
-        "precision": precision,
-        "weight_type": weight_type,
-        "status": "PENDING",
-        "submitted_time": current_time,
-        "model_type": model_type,
-        "likes": model_info.likes,
-        "params": model_size,
-        "license": license,
-        "private": False,
-    }
-
-    # Check for duplicate submission
-    if f"{model}_{revision}_{precision}" in REQUESTED_MODELS:
-        return styled_warning("This model has been already submitted.")
-
-    print("Creating eval file")
-    OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
-    os.makedirs(OUT_DIR, exist_ok=True)
-    out_path = f"{OUT_DIR}/{model_path}_eval_request_False_{precision}_{weight_type}.json"
-
-    with open(out_path, "w") as f:
-        f.write(json.dumps(eval_entry))
-
-    print("Uploading eval file")
-    API.upload_file(
-        path_or_fileobj=out_path,
-        path_in_repo=out_path.split("eval-queue/")[1],
-        repo_id=QUEUE_REPO,
-        repo_type="dataset",
-        commit_message=f"Add {model} to eval queue",
-    )
-
-    # Remove the local file
-    os.remove(out_path)
-
-    return styled_message(
-        "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list."
-    )