Spaces:
Sleeping
Sleeping
| import gradio as gr | |
| import plotly.graph_objects as go | |
| import os | |
| # ============ Leaderboard Data ============ | |
| MODEL_EVALS = { | |
| "LLM (General OSIR)": { | |
| "Nexa Mistral Sci-7B": 0.61, | |
| "Llama-3-8B-Instruct": 0.39, | |
| "Mixtral-8x7B-Instruct-v0.1": 0.41, | |
| "Claude-3-Sonnet": 0.64, | |
| "GPT-4-Turbo": 0.68, | |
| "GPT-4o": 0.71, | |
| }, | |
| "LLM (Field-Specific OSIR)": { | |
| "Nexa Bio Adapter": 0.66, | |
| "Nexa Astro Adapter": 0.70, | |
| "GPT-4o (Biomed)": 0.69, | |
| "Claude-3-Opus (Bio)": 0.67, | |
| "Llama-3-8B-Bio": 0.42, | |
| "Mixtral-8x7B-BioTune": 0.43, | |
| }, | |
| } | |
| # ============ Plotting Function ============ | |
| def plot_domain(domain): | |
| sorted_items = sorted(MODEL_EVALS[domain].items(), key=lambda x: x[1], reverse=True) | |
| models, scores = zip(*sorted_items) | |
| fig = go.Figure() | |
| fig.add_trace(go.Bar( | |
| x=scores, | |
| y=models, | |
| orientation='h', | |
| marker_color='lightblue', | |
| )) | |
| fig.update_layout( | |
| title=f"Model vs. Overall Score — {domain}", | |
| xaxis_title="Scientific Utility Score", | |
| yaxis_title="Model", | |
| xaxis_range=[0, 1.0], | |
| template="plotly_white", | |
| height=500, | |
| margin=dict(l=120, r=20, t=40, b=40), | |
| ) | |
| return fig | |
| # ============ Upload Handling (for later use) ============ | |
| def handle_upload(file): | |
| if file is not None: | |
| return f"Uploaded: {file.name}" | |
| return "No file uploaded." | |
| # ============ Gradio UI ============ | |
| with gr.Blocks(css="body {font-family: 'Inter', sans-serif; background-color: #fafafa;}") as demo: | |
| gr.Markdown(""" | |
| # 🧠 SciEval | OSIR Leaderboard | |
| Welcome to the **OSIR** benchmarking suite for evaluating scientific language models. This leaderboard tracks general-purpose and field-specific reasoning performance across the **SciEval** benchmark. | |
| """) | |
| with gr.Row(): | |
| with gr.Column(): | |
| domain_choice = gr.Dropdown(choices=list(MODEL_EVALS.keys()), label="Select Evaluation Domain", value="LLM (General OSIR)") | |
| leaderboard_plot = gr.Plot() | |
| domain_choice.change(fn=plot_domain, inputs=domain_choice, outputs=leaderboard_plot) | |
| with gr.Column(): | |
| gr.Markdown(""" | |
| ### 📄 Upload Model Output | |
| Upload a generated scientific paper or abstract (PDF or TXT). | |
| """) | |
| upload = gr.File(file_types=[".pdf", ".txt"]) | |
| upload_btn = gr.Button("Submit File") | |
| result = gr.Textbox(label="Upload Status") | |
| upload_btn.click(fn=handle_upload, inputs=upload, outputs=result) | |
| gr.Markdown(""" | |
| --- | |
| ### ℹ️ About | |
| **SciEval** is a model-agnostic benchmark to evaluate the scientific utility of LLMs under the **OSIR** (Open Scientific Intelligence & Reasoning) initiative. We score models based on: | |
| - Information entropy & novelty | |
| - Internal consistency | |
| - Hypothesis framing | |
| - Domain grounding & math logic | |
| - Scientific utility (overall use to researchers) | |
| This leaderboard includes Nexa's adapters and comparisons to general-purpose LLMs like GPT-4o, Claude 3, and open-source Mistral / LLaMA. | |
| """) | |
| leaderboard_plot.render() | |
| demo.launch() | |