Upload 2 files

app.py

@@ -1,15 +1,19 @@
 import os.path
+
 import gradio as gr
+import numpy as np
 import pandas as pd
+
 from constants import *
 
-# ------------ 下载链接 ------------
+
 def get_download_link_model(task, dataset, example):
     _task_path = TASK_PATH_MAPPING[task]
     _dataset_path = DATASET_PATH_MAPPING[dataset]
     _example_path = EXAMPLE_PATH_MAPPING[example]
     return os.path.join("data", _task_path, _dataset_path, "weight", f"{_example_path}.zip")
 
+
 def get_download_link_json(task, dataset, example):
     _task_path = TASK_PATH_MAPPING[task]
     _dataset_path = DATASET_PATH_MAPPING[dataset]
@@ -19,48 +23,48 @@ def get_download_link_json(task, dataset, example):
     else:
         return os.path.join("data", _task_path, _dataset_path, "json", f"{_example_path}.json")
 
-# ------------ 数据读取 + 平均准确率 ------------
+
 def get_data(task, dataset, example):
     _task_path = TASK_PATH_MAPPING[task]
     _dataset_path = DATASET_PATH_MAPPING[dataset]
     _example_path = EXAMPLE_PATH_MAPPING[example]
     csv_file = os.path.join("data", _task_path, _dataset_path, "csv", f"{_example_path}.csv")
     if not os.path.exists(csv_file):
-        return None, None
-
+        return
     read_data = pd.read_csv(csv_file)
     data = pd.DataFrame(columns=COLUMN_NAMES)
-    average_acc = None
 
     if _task_path == "coding":
-        for _, row in read_data.iterrows():
-            data = pd.concat([data, pd.DataFrame([{
+        for index, row in read_data.iterrows():
+            data = data._append({
                 "Prompt": row["prompt"],
                 "Pass@1": round(float(row["pass@1"]) * 100, 3),
                 "Pass@5": round(float(row["pass@5"]) * 100, 3),
                 "Pass@10": round(float(row["pass@10"]) * 100, 3),
                 "Correctness": "N/A"
-            }])], ignore_index=True)
-        # 仅对 HumanEval 数据集计算三列平均
-        if "HumanEval" in dataset:
-            p1_mean = round(read_data["pass@1"].mean() * 100, 3)
-            p5_mean = round(read_data["pass@5"].mean() * 100, 3)
-            p10_mean = round(read_data["pass@10"].mean() * 100, 3)
-            average_acc = f"{p1_mean} / {p5_mean} / {p10_mean}"
-    elif _task_path in ["common", "math"]:
-        for _, row in read_data.iterrows():
-            data = pd.concat([data, pd.DataFrame([{
+            }, ignore_index=True)
+    elif _task_path == "common":
+        for index, row in read_data.iterrows():
+            data = data._append({
+                "Prompt": row["prompt"],
+                "Pass@1": None,
+                "Pass@5": None,
+                "Pass@10": None,
+                "Correctness": "✅" if row["correctness"] else "❌"
+            }, ignore_index=True)
+    elif _task_path == "math":
+        for index, row in read_data.iterrows():
+            data = data._append({
                 "Prompt": row["prompt"],
                 "Pass@1": None,
                 "Pass@5": None,
                 "Pass@10": None,
                 "Correctness": "✅" if row["correctness"] else "❌"
-            }])], ignore_index=True)
-        average_acc = round(read_data["correctness"].mean() * 100, 3)
+            }, ignore_index=True)
+
+    return data
 
-    return data, average_acc
 
-# ------------ Gradio UI ------------
 with gr.Blocks() as demo_board:
     gr.HTML(DND_HEADER)
     gr.Markdown(DND_INTRODUCTION)
@@ -84,20 +88,8 @@ with gr.Blocks() as demo_board:
         interactive=True,
     )
 
-    # 平均准确率（放在 Prompt 表格上方）
-    average_acc_display = gr.Textbox(
-        label="Average Accuracy (%)",
-        value=lambda: str(get_data(task.value, dataset.value, example.value)[1]),
-        interactive=False,
-        visible=True,
-        scale=0,
-        max_lines=1,
-        min_width=160
-    )
-
-    # Prompt 表格
     board = gr.components.Dataframe(
-        value=lambda: get_data(task.value, dataset.value, example.value)[0],
+        value=get_data(task.value, dataset.value, example.value),
         column_widths=["60%", "10%", "10%", "10%", "10%"],
         headers=COLUMN_NAMES,
         type="pandas",
@@ -107,31 +99,28 @@ with gr.Blocks() as demo_board:
         max_height=500,
     )
 
-    # 联动更新：task -> dataset
-    task.change(
-        lambda t: gr.Radio(
-            label="Dataset",
-            choices=TASK_DATASET_LIST[t],
-            value=TASK_DATASET_LIST[t][0],
-            interactive=True,
-        ),
-        inputs=[task],
-        outputs=dataset
-    )
+    task.change(lambda t: gr.Radio(
+        label="Dataset",
+        choices=TASK_DATASET_LIST[t],
+        value=TASK_DATASET_LIST[t][0],
+        interactive=True,
+    ), inputs=[task], outputs=dataset)
 
-    # 联动更新：task / dataset / example -> 表格 + 平均准确率
     for component in [task, dataset, example]:
-        component.change(
-            lambda t, d, e: (get_data(t, d, e)[0], str(get_data(t, d, e)[1])),
-            inputs=[task, dataset, example],
-            outputs=[board, average_acc_display]
-        )
+        component.change(lambda t, d, e: gr.components.Dataframe(
+            value=get_data(t, d, e),
+            column_widths=["60%", "10%", "10%", "10%", "10%"],
+            headers=COLUMN_NAMES,
+            type="pandas",
+            datatype=DATA_TITLE_TYPE,
+            interactive=False,
+            visible=True,
+            max_height=500,
+        ), inputs=[task, dataset, example], outputs=board)
 
-    # 下载按钮
     with gr.Row():
         json_downloader = gr.DownloadButton("Download JSON", visible=True)
         model_downloader = gr.DownloadButton("Download Model", visible=True)
-
         json_downloader.click(
             fn=get_download_link_json,
             inputs=[task, dataset, example],
@@ -143,7 +132,6 @@ with gr.Blocks() as demo_board:
             outputs=model_downloader,
         )
 
-    # 引用文本
     citation_button = gr.Textbox(
         value=CITATION_BUTTON_TEXT,
         label=CITATION_BUTTON_LABEL,
@@ -152,5 +140,4 @@ with gr.Blocks() as demo_board:
         show_copy_button=True,
     )
 
-# 启动
-demo_board.launch()
+demo_board.launch()

constants.py

@@ -0,0 +1,81 @@
+DND_HEADER = """
+<style>
+.header-gradient {
+    top: 40%;
+    bottom: 40%;
+    padding: 10px 0px;
+    font-weight: bold;
+    font-size: 40px;
+    font-family: Inter, Arial, Helvetica, sans-serif;
+    background: linear-gradient(to right, #67a102, #c0dc90);
+    -webkit-text-fill-color: transparent;
+    -webkit-background-clip: text;
+}
+
+.header-normal {
+    top: 40%;
+    bottom: 40%;
+    padding: 10px 0px;
+    font-weight: bold;
+    font-size: 40px;
+    font-family: Inter, Arial, Helvetica, sans-serif;
+}
+</style>
+
+<div align="center">
+    <span class="header-gradient"> Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights </span>
+</div>
+<p align="center">
+| <a href=""><b>Documentation</b></a> | <a href=""><b>Github</b></a> | <a href="https://arxiv.org/abs/2506.16406"><b>Paper </b> </a> | <a href="https://x.com/VictorKaiWang1/status/1935905121659240513"><b>Twitter/X</b> </a> |
+</p>"""
+
+DND_INTRODUCTION = """
+🚀 Welcome to the Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights! 
+
+> Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights is a zero-shot prompt-to-weights model that can generate a model from a prompt.
+
+- **Zero-Shot**: Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights can generate a model from a prompt without any training data.
+- **Prompt-to-Weights**: Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights can generate a model from a prompt.
+- **Easy-to-use**: Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights provides a unified interface for prompt-to-weights model generation.
+
+"""
+
+TASK_LIST = ["🧠 Commonsense Reasoning", "🔢 Math", "💻 Coding"]
+TASK_DATASET_LIST = {
+    "🧠 Commonsense Reasoning": ["ARC-c", "OBQA"],
+    "🔢 Math": ["GSM-8K"],
+    "💻 Coding": ["HumanEval"],
+}
+EXAMPLE_LIST = ["Example 1", "Example 2", "Example 3", "Example 4", "Example 5"]
+
+TASK_PATH_MAPPING = {
+    "🧠 Commonsense Reasoning": "common",
+    "🔢 Math": "math",
+    "💻 Coding": "coding",
+}
+DATASET_PATH_MAPPING = {
+    "ARC-c": "arc_c",
+    "OBQA": "obqa",
+    "GSM-8K": "gsm8k",
+    "HumanEval": "humaneval",
+}
+EXAMPLE_PATH_MAPPING = {
+    "Example 1": "1",
+    "Example 2": "2",
+    "Example 3": "3",
+    "Example 4": "4",
+    "Example 5": "5",
+}
+
+COLUMN_NAMES = ["Prompt", "Pass@1", "Pass@5", "Pass@10", "Correctness"]
+DATA_TITLE_TYPE = ['markdown', 'number', 'number', 'number', 'markdown']
+
+CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
+CITATION_BUTTON_TEXT = r"""
+@article{liang2025drag,
+  title={Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights},
+  author={Liang, Zhiyuan and Tang, Dongwen and Zhou, Yuhao and Zhao, Xuanlei and Shi, Mingjia and Zhao, Wangbo and Li, Zekai and Wang, Peihao and Sch{\"u}rholt, Konstantin and Borth, Damian and others},
+  journal={arXiv preprint arXiv:2506.16406},
+  year={2025}
+}
+"""