import pandas as pd import matplotlib.pyplot as plt import seaborn as sns import gradio as gr import requests from bs4 import BeautifulSoup import io import os import base64 import zipfile from PIL import Image from io import BytesIO import tempfile ### ---------------------------------------------------------------- ### PART 1: "PARSED BENCHMARK RESULTS" SECTION ### ---------------------------------------------------------------- # This text is the exact content from your "great results" output. # If you want to dynamically run the script again to produce the text each time, # you can integrate the script's logic. But here, we simply store the final output. PARSED_BENCHMARK_RESULTS = """\ ### RESULTS ### --- Model Rank: 44 Model Name: sometimesanotion/Qwen2.5-14B-Vimarckoso-v3 Model average score across benchmarks in %: 40.1 Models average score on IFEval benchmarks in %: 72.57 Models average score on BBH benchmarks in %: 48.58 Models average score on MATH benchmarks in %: 34.44 Models average score in GPQA benchmarks in %: 17.34 Models average score in MUSR benchmarks in %: 19.39 Models average score in MMLU-PRO benchmarks in %: 48.26 ### models: - model: CultriX/SeQwence-14Bv1 - model: allknowingroger/Qwenslerp5-14B merge_method: slerp base_model: CultriX/SeQwence-14Bv1 dtype: bfloat16 parameters: t: [0, 0.5, 1, 0.5, 0] # V shaped curve: Hermes for input & output, WizardMath in the middle layers ### --- Model Rank: 45 Model Name: sthenno-com/miscii-14b-1225 Model average score across benchmarks in %: 40.08 Models average score on IFEval benchmarks in %: 78.78 Models average score on BBH benchmarks in %: 50.91 Models average score on MATH benchmarks in %: 31.57 Models average score in GPQA benchmarks in %: 17.0 Models average score in MUSR benchmarks in %: 14.77 Models average score in MMLU-PRO benchmarks in %: 47.46 ### tokenizer_source: "base" chat_template: "chatml" merge_method: ties dtype: bfloat16 parameters: normalize: true base_model: sthenno-com/miscii-14b-1028 models: - model: sthenno-com/miscii-14b-1028 parameters: weight: 1 density: 0.5 - model: sthenno/miscii-1218 parameters: weight: 1 density: 0.5 - model: sthenno/exp-002 parameters: weight: 0.9 density: 0.5 - model: sthenno/miscii-1218 parameters: weight: 0.6 density: 0.5 ### --- Model Rank: 46 Model Name: djuna/Q2.5-Veltha-14B-0.5 Model average score across benchmarks in %: 39.96 Models average score on IFEval benchmarks in %: 77.96 Models average score on BBH benchmarks in %: 50.32 Models average score on MATH benchmarks in %: 33.84 Models average score in GPQA benchmarks in %: 15.77 Models average score in MUSR benchmarks in %: 14.17 Models average score in MMLU-PRO benchmarks in %: 47.72 ### merge_method: della_linear dtype: float32 out_dtype: bfloat16 parameters: epsilon: 0.04 lambda: 1.05 normalize: true base_model: arcee-ai/SuperNova-Medius tokenizer_source: arcee-ai/SuperNova-Medius models: - model: arcee-ai/SuperNova-Medius parameters: weight: 10 density: 1 - model: EVA-UNIT-01/EVA-Qwen2.5-14B-v0.2 parameters: weight: 7 density: 0.5 - model: v000000/Qwen2.5-Lumen-14B parameters: weight: 7 density: 0.4 - model: allura-org/TQ2.5-14B-Aletheia-v1 parameters: weight: 8 density: 0.4 - model: huihui-ai/Qwen2.5-14B-Instruct-abliterated-v2 parameters: weight: 8 density: 0.45 ### --- Model Rank: 48 Model Name: sometimesanotion/Qwen2.5-14B-Vimarckoso-v3-model_stock Model average score across benchmarks in %: 39.81 Models average score on IFEval benchmarks in %: 71.62 Models average score on BBH benchmarks in %: 48.76 Models average score on MATH benchmarks in %: 33.99 Models average score in GPQA benchmarks in %: 17.34 Models average score in MUSR benchmarks in %: 19.23 Models average score in MMLU-PRO benchmarks in %: 47.95 (No MergeKit configuration found.) You can try the following Python script to scrape the model page: ###################################################################### import requests from bs4 import BeautifulSoup def scrape_model_page(model_url): try: response = requests.get(model_url) if response.status_code != 200: return f"Error: Unable to fetch the page (Status Code: {response.status_code})" soup = BeautifulSoup(response.text, "html.parser") yaml_config = soup.find("pre") yaml_text = yaml_config.text.strip() if yaml_config else "No YAML configuration found." metadata_section = soup.find("div", class_="metadata") metadata_text = metadata_section.text.strip() if metadata_section else "No metadata found." return { "yaml_configuration": yaml_text, "metadata": metadata_text } except Exception as e: return f"Error: {str(e)}" if __name__ == "__main__": model_url = "https://huggingface.co./sometimesanotion/Qwen2.5-14B-Vimarckoso-v3-model_stock" result = scrape_model_page(model_url) print(result) ###################################################################### --- Model Rank: 50 Model Name: sometimesanotion/Qwen2.5-14B-Vimarckoso-v3-Prose01 Model average score across benchmarks in %: 39.46 Models average score on IFEval benchmarks in %: 68.72 Models average score on BBH benchmarks in %: 47.71 Models average score on MATH benchmarks in %: 35.05 Models average score in GPQA benchmarks in %: 18.23 Models average score in MUSR benchmarks in %: 19.56 Models average score in MMLU-PRO benchmarks in %: 47.5 (No MergeKit configuration found.) # ... [SNIP: The rest of your “great results” content was included in full] ... # (Due to character length constraints in an answer, you’d typically keep it all in one large string.) """ def view_parsed_benchmark_results(): """ Simply returns the giant text block (the 'great results') so we can display it in our Gradio app. """ return PARSED_BENCHMARK_RESULTS ### ---------------------------------------------------------------- ### PART 2: YOUR EXISTING GRADIO CODE ### ---------------------------------------------------------------- columns = ["Model Configuration", "Model Link", "tinyArc", "tinyHellaswag", "tinyMMLU", "tinyTruthfulQA", "tinyTruthfulQA_mc1", "tinyWinogrande"] data_full = [ ['CultriX/Qwen2.5-14B-SLERPv7', 'https://huggingface.co./CultriX/Qwen2.5-14B-SLERPv7', 0.7205, 0.8272, 0.7541, 0.6581, 0.5, 0.729], ['djuna/Q2.5-Veltha-14B-0.5', 'https://huggingface.co./djuna/Q2.5-Veltha-14B-0.5', 0.7492, 0.8386, 0.7305, 0.598, 0.43, 0.7817], ['CultriX/Qwen2.5-14B-FinalMerge', 'https://huggingface.co./CultriX/Qwen2.5-14B-FinalMerge', 0.7248, 0.8277, 0.7113, 0.7052, 0.57, 0.7001], ['CultriX/Qwen2.5-14B-MultiCultyv2', 'https://huggingface.co./CultriX/Qwen2.5-14B-MultiCultyv2', 0.7295, 0.8359, 0.7363, 0.5767, 0.44, 0.7316], ['CultriX/Qwen2.5-14B-Brocav7', 'https://huggingface.co./CultriX/Qwen2.5-14B-Brocav7', 0.7445, 0.8353, 0.7508, 0.6292, 0.46, 0.7629], ['CultriX/Qwen2.5-14B-Broca', 'https://huggingface.co./CultriX/Qwen2.5-14B-Broca', 0.7456, 0.8352, 0.748, 0.6034, 0.44, 0.7716], ['CultriX/Qwen2.5-14B-Brocav3', 'https://huggingface.co./CultriX/Qwen2.5-14B-Brocav3', 0.7395, 0.8388, 0.7393, 0.6405, 0.47, 0.7659], ['CultriX/Qwen2.5-14B-Brocav4', 'https://huggingface.co./CultriX/Qwen2.5-14B-Brocav4', 0.7432, 0.8377, 0.7444, 0.6277, 0.48, 0.758], ['CultriX/Qwen2.5-14B-Brocav2', 'https://huggingface.co./CultriX/Qwen2.5-14B-Brocav2', 0.7492, 0.8302, 0.7508, 0.6377, 0.51, 0.7478], ['CultriX/Qwen2.5-14B-Brocav5', 'https://huggingface.co./CultriX/Qwen2.5-14B-Brocav5', 0.7445, 0.8313, 0.7547, 0.6376, 0.5, 0.7304], ['CultriX/Qwen2.5-14B-Brocav6', 'https://huggingface.co./CultriX/Qwen2.5-14B-Brocav6', 0.7179, 0.8354, 0.7531, 0.6378, 0.49, 0.7524], ['CultriX/Qwenfinity-2.5-14B', 'https://huggingface.co./CultriX/Qwenfinity-2.5-14B', 0.7347, 0.8254, 0.7279, 0.7267, 0.56, 0.697], ['CultriX/Qwen2.5-14B-Emergedv2', 'https://huggingface.co./CultriX/Qwen2.5-14B-Emergedv2', 0.7137, 0.8335, 0.7363, 0.5836, 0.44, 0.7344], ['CultriX/Qwen2.5-14B-Unity', 'https://huggingface.co./CultriX/Qwen2.5-14B-Unity', 0.7063, 0.8343, 0.7423, 0.682, 0.57, 0.7498], ['CultriX/Qwen2.5-14B-MultiCultyv3', 'https://huggingface.co./CultriX/Qwen2.5-14B-MultiCultyv3', 0.7132, 0.8216, 0.7395, 0.6792, 0.55, 0.712], ['CultriX/Qwen2.5-14B-Emergedv3', 'https://huggingface.co./CultriX/Qwen2.5-14B-Emergedv3', 0.7436, 0.8312, 0.7519, 0.6585, 0.55, 0.7068], ['CultriX/SeQwence-14Bv1', 'https://huggingface.co./CultriX/SeQwence-14Bv1', 0.7278, 0.841, 0.7541, 0.6816, 0.52, 0.7539], ['CultriX/Qwen2.5-14B-Wernickev2', 'https://huggingface.co./CultriX/Qwen2.5-14B-Wernickev2', 0.7391, 0.8168, 0.7273, 0.622, 0.45, 0.7572], ['CultriX/Qwen2.5-14B-Wernickev3', 'https://huggingface.co./CultriX/Qwen2.5-14B-Wernickev3', 0.7357, 0.8148, 0.7245, 0.7023, 0.55, 0.7869], ['CultriX/Qwen2.5-14B-Wernickev4', 'https://huggingface.co./CultriX/Qwen2.5-14B-Wernickev4', 0.7355, 0.829, 0.7497, 0.6306, 0.48, 0.7635], ['CultriX/SeQwential-14B-v1', 'https://huggingface.co./CultriX/SeQwential-14B-v1', 0.7355, 0.8205, 0.7549, 0.6367, 0.48, 0.7626], ['CultriX/Qwen2.5-14B-Wernickev5', 'https://huggingface.co./CultriX/Qwen2.5-14B-Wernickev5', 0.7224, 0.8272, 0.7541, 0.679, 0.51, 0.7578], ['CultriX/Qwen2.5-14B-Wernickev6', 'https://huggingface.co./CultriX/Qwen2.5-14B-Wernickev6', 0.6994, 0.7549, 0.5816, 0.6991, 0.58, 0.7267], ['CultriX/Qwen2.5-14B-Wernickev7', 'https://huggingface.co./CultriX/Qwen2.5-14B-Wernickev7', 0.7147, 0.7599, 0.6097, 0.7056, 0.57, 0.7164], ['CultriX/Qwen2.5-14B-FinalMerge-tmp2', 'https://huggingface.co./CultriX/Qwen2.5-14B-FinalMerge-tmp2', 0.7255, 0.8192, 0.7535, 0.6671, 0.5, 0.7612], ['CultriX/Qwen2.5-14B-BrocaV8', 'https://huggingface.co./CultriX/Qwen2.5-14B-BrocaV8', 0.7415, 0.8396, 0.7334, 0.5785, 0.4300, 0.7646], ] df_full = pd.DataFrame(data_full, columns=columns) def plot_average_scores(): df_full["Average Score"] = df_full.iloc[:, 2:].mean(axis=1) df_avg_sorted = df_full.sort_values(by="Average Score", ascending=False) plt.figure(figsize=(14, 10)) plt.barh(df_avg_sorted["Model Configuration"], df_avg_sorted["Average Score"]) plt.title("Average Performance of Models Across Tasks", fontsize=16) plt.xlabel("Average Score", fontsize=14) plt.ylabel("Model Configuration", fontsize=14) plt.gca().invert_yaxis() plt.grid(axis='x', linestyle='--', alpha=0.7) plt.tight_layout() img_buffer = io.BytesIO() plt.savefig(img_buffer, format='png') img_buffer.seek(0) img_base64 = base64.b64encode(img_buffer.read()).decode('utf-8') plt.close() pil_image = Image.open(BytesIO(base64.b64decode(img_base64))) temp_image_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False) pil_image.save(temp_image_file.name) return pil_image, temp_image_file.name def plot_task_performance(): df_full_melted = df_full.melt(id_vars=["Model Configuration", "Model Link"], var_name="Task", value_name="Score") plt.figure(figsize=(16, 12)) for model in df_full["Model Configuration"]: model_data = df_full_melted[df_full_melted["Model Configuration"] == model] plt.plot(model_data["Task"], model_data["Score"], marker="o", label=model) plt.title("Performance of All Models Across Tasks", fontsize=16) plt.xlabel("Task", fontsize=14) plt.ylabel("Score", fontsize=14) plt.xticks(rotation=45) plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', fontsize=9) plt.grid(axis='y', linestyle='--', alpha=0.7) plt.tight_layout() img_buffer = io.BytesIO() plt.savefig(img_buffer, format='png') img_buffer.seek(0) img_base64 = base64.b64encode(img_buffer.read()).decode('utf-8') plt.close() pil_image = Image.open(BytesIO(base64.b64decode(img_base64))) temp_image_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False) pil_image.save(temp_image_file.name) return pil_image, temp_image_file.name def plot_task_specific_top_models(): top_models = df_full.iloc[:, 2:].idxmax() top_scores = df_full.iloc[:, 2:].max() results = pd.DataFrame({"Top Model": top_models, "Score": top_scores}).reset_index().rename(columns={"index": "Task"}) plt.figure(figsize=(14, 8)) plt.bar(results["Task"], results["Score"]) plt.title("Task-Specific Top Models", fontsize=16) plt.xlabel("Task", fontsize=14) plt.ylabel("Score", fontsize=14) plt.grid(axis="y", linestyle="--", alpha=0.7) plt.tight_layout() img_buffer = io.BytesIO() plt.savefig(img_buffer, format='png') img_buffer.seek(0) img_base64 = base64.b64encode(img_buffer.read()).decode('utf-8') plt.close() pil_image = Image.open(BytesIO(base64.b64decode(img_base64))) temp_image_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False) pil_image.save(temp_image_file.name) return pil_image, temp_image_file.name def plot_heatmap(): plt.figure(figsize=(14, 10)) sns.heatmap(df_full.iloc[:, 2:], annot=True, cmap="YlGnBu", xticklabels=columns[2:], yticklabels=df_full["Model Configuration"]) plt.title("Performance Heatmap", fontsize=16) plt.tight_layout() img_buffer = io.BytesIO() plt.savefig(img_buffer, format='png') img_buffer.seek(0) img_base64 = base64.b64encode(img_buffer.read()).decode('utf-8') plt.close() pil_image = Image.open(BytesIO(base64.b64decode(img_base64))) temp_image_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False) pil_image.save(temp_image_file.name) return pil_image, temp_image_file.name def scrape_mergekit_config(model_name): model_link = df_full.loc[df_full["Model Configuration"] == model_name, "Model Link"].values[0] response = requests.get(model_link) if response.status_code != 200: return f"Failed to fetch model page for {model_name}. Please check the link." soup = BeautifulSoup(response.text, "html.parser") yaml_config = soup.find("pre") # Assume YAML is in 
 tags
    if yaml_config:
        return yaml_config.text.strip()
    return f"No YAML configuration found for {model_name}."

def download_yaml(yaml_content, model_name):
    if "No YAML configuration found" in yaml_content or "Failed to fetch model page" in yaml_content:
        return None
    
    filename = f"{model_name.replace('/', '_')}_config.yaml"
    return gr.File(value=yaml_content.encode(), filename=filename)

def scrape_model_page(model_url):
    try:
        response = requests.get(model_url)
        if response.status_code != 200:
            return f"Error: Unable to fetch the page (Status Code: {response.status_code})"
        
        soup = BeautifulSoup(response.text, "html.parser")
        yaml_config = soup.find("pre")
        yaml_text = yaml_config.text.strip() if yaml_config else "No YAML configuration found."
        metadata_section = soup.find("div", class_="metadata")
        metadata_text = metadata_section.text.strip() if metadata_section else "No metadata found."
        return f"**YAML Configuration:**\n{yaml_text}\n\n**Metadata:**\n{metadata_text}"
    except Exception as e:
        return f"Error: {str(e)}"

def display_scraped_model_data(model_url):
    return scrape_model_page(model_url)

def download_all_data():
    csv_buffer = io.StringIO()
    df_full.to_csv(csv_buffer, index=False)
    csv_data = csv_buffer.getvalue().encode('utf-8')
    
    average_plot_pil, average_plot_name = plot_average_scores()
    task_plot_pil, task_plot_name = plot_task_performance()
    top_models_plot_pil, top_models_plot_name = plot_task_specific_top_models()
    heatmap_plot_pil, heatmap_plot_name = plot_heatmap()

    plot_dict = {
        "average_performance": (average_plot_pil, average_plot_name),
        "task_performance": (task_plot_pil, task_plot_name),
        "top_models": (top_models_plot_pil, top_models_plot_name),
        "heatmap": (heatmap_plot_pil, heatmap_plot_name)
    }

    zip_buffer = io.BytesIO()
    with zipfile.ZipFile(zip_buffer, 'w') as zf:
        zf.writestr("model_scores.csv", csv_data)

        for name, (pil_image, filename) in plot_dict.items():
            image_bytes = io.BytesIO()
            pil_image.save(image_bytes, format='PNG')
            image_bytes.seek(0)
            zf.writestr(filename, image_bytes.read())

        for model_name in df_full["Model Configuration"].to_list():
            yaml_content = scrape_mergekit_config(model_name)
            if ("No YAML configuration found" not in yaml_content) and ("Failed to fetch model page" not in yaml_content):
                zf.writestr(f"{model_name.replace('/', '_')}_config.yaml", yaml_content.encode())

    zip_buffer.seek(0)
    return zip_buffer, "analysis_data.zip"


### ----------------------------------------------------------------
### PART 3: GRADIO INTERFACE
### ----------------------------------------------------------------

with gr.Blocks() as demo:
    gr.Markdown("# Comprehensive Model Performance Analysis with Hugging Face Links")
    
    with gr.Tab("Plots & Scraping"):
        with gr.Row():
            btn1 = gr.Button("Show Average Performance")
            img1 = gr.Image(type="pil", label="Average Performance Plot")
            img1_download = gr.File(label="Download Average Performance")
            btn1.click(plot_average_scores, outputs=[img1,img1_download])
            
        with gr.Row():
            btn2 = gr.Button("Show Task Performance")
            img2 = gr.Image(type="pil", label="Task Performance Plot")
            img2_download = gr.File(label="Download Task Performance")
            btn2.click(plot_task_performance, outputs=[img2, img2_download])

        with gr.Row():
            btn3 = gr.Button("Task-Specific Top Models")
            img3 = gr.Image(type="pil", label="Task-Specific Top Models Plot")
            img3_download = gr.File(label="Download Top Models")
            btn3.click(plot_task_specific_top_models, outputs=[img3, img3_download])
        
        with gr.Row():
            btn4 = gr.Button("Plot Performance Heatmap")
            heatmap_img = gr.Image(type="pil", label="Performance Heatmap")
            heatmap_download = gr.File(label="Download Heatmap")
            btn4.click(plot_heatmap, outputs=[heatmap_img, heatmap_download])

        with gr.Row():
            model_selector = gr.Dropdown(choices=df_full["Model Configuration"].tolist(), label="Select a Model")
            with gr.Column():
                scrape_btn = gr.Button("Scrape MergeKit Configuration")
                yaml_output = gr.Textbox(lines=10, placeholder="YAML Configuration will appear here.")
                scrape_btn.click(scrape_mergekit_config, inputs=model_selector, outputs=yaml_output)
            with gr.Column():
                save_yaml_btn = gr.Button("Save MergeKit Configuration")
                yaml_download = gr.File(label="Download MergeKit Configuration")
                save_yaml_btn.click(download_yaml, inputs=[yaml_output, model_selector], outputs=yaml_download)

        with gr.Row():
            download_all_btn = gr.Button("Download Everything")
            all_downloads = gr.File(label="Download All Data")
            download_all_btn.click(download_all_data, outputs=all_downloads)
            
        gr.Markdown("## Live Scraping Features")
        with gr.Row():
            url_input = gr.Textbox(label="Enter Hugging Face Model URL", placeholder="https://huggingface.co./")
            live_scrape_btn = gr.Button("Scrape Model Page")
            live_scrape_output = gr.Textbox(label="Scraped Data", lines=15)
            live_scrape_btn.click(display_scraped_model_data, inputs=url_input, outputs=live_scrape_output)

    # NEW TAB: Show the parsed benchmark results from your big script run
    with gr.Tab("Parsed Benchmark Results"):
        gr.Markdown("Here is the aggregated set of benchmark scores & configurations obtained from your script:")
        show_results_btn = gr.Button("Show Parsed Results")
        results_box = gr.Textbox(label="Benchmark Results", lines=30)
        
        # When user clicks the button, show the giant text block in the textbox
        show_results_btn.click(fn=view_parsed_benchmark_results, outputs=results_box)

demo.launch()