app.py

from pathlib import Path

import gradio as gr
import pandas as pd
import spaces
from huggingface_hub import hf_hub_download
from llama_cpp import Llama


MODEL_REPO = "mradermacher/Spreadsheet-RL-4B-GGUF"
QUANT_FILES = {
    "Q2_K · 1.9 GB": "Spreadsheet-RL-4B.Q2_K.gguf",
    "Q3_K_S · 2.2 GB": "Spreadsheet-RL-4B.Q3_K_S.gguf",
    "Q3_K_M · 2.3 GB · lower quality": "Spreadsheet-RL-4B.Q3_K_M.gguf",
    "Q3_K_L · 2.5 GB": "Spreadsheet-RL-4B.Q3_K_L.gguf",
    "IQ4_XS · 2.6 GB": "Spreadsheet-RL-4B.IQ4_XS.gguf",
    "Q4_K_S · 2.7 GB · recommended": "Spreadsheet-RL-4B.Q4_K_S.gguf",
    "Q4_K_M · 2.8 GB · recommended": "Spreadsheet-RL-4B.Q4_K_M.gguf",
    "Q5_K_S · 3.2 GB": "Spreadsheet-RL-4B.Q5_K_S.gguf",
    "Q5_K_M · 3.3 GB": "Spreadsheet-RL-4B.Q5_K_M.gguf",
    "Q6_K · 3.7 GB · very good quality": "Spreadsheet-RL-4B.Q6_K.gguf",
    "Q8_0 · 4.8 GB · best quality": "Spreadsheet-RL-4B.Q8_0.gguf",
    "f16 · 8.9 GB": "Spreadsheet-RL-4B.f16.gguf",
}

model = None
active_quant = None


def download_quant(quantization: str) -> None:
    hf_hub_download(repo_id=MODEL_REPO, filename=QUANT_FILES[quantization])


def file_to_text(file_path: str | None) -> str:
    if file_path is None:
        return ""

    path = Path(file_path)
    suffix = path.suffix.lower()

    if suffix in {".xlsx", ".xls"}:
        sheets = pd.read_excel(path, sheet_name=None)
        return "\n\n".join(
            f"## Sheet: {sheet_name}\n{frame.to_csv(index=False)}"
            for sheet_name, frame in sheets.items()
        )

    if suffix == ".csv":
        return pd.read_csv(path).to_csv(index=False)

    if suffix == ".tsv":
        return pd.read_csv(path, sep="\t").to_csv(index=False)

    return path.read_text(encoding="utf-8")


@spaces.GPU(duration=120)
def generate(
    system_prompt: str,
    user_prompt: str,
    attachment: str | None,
    quantization: str,
) -> str:
    global active_quant, model

    quant_file = QUANT_FILES[quantization]
    if active_quant != quantization:
        model = None
        active_quant = None
        model = Llama(
            model_path=hf_hub_download(repo_id=MODEL_REPO, filename=quant_file),
            n_ctx=4096,
            n_gpu_layers=-1,
            verbose=True,
        )
        active_quant = quantization

    attachment_text = file_to_text(attachment)
    user_content = user_prompt
    if attachment_text:
        user_content = f"{user_prompt}\n\n<attachment>\n{attachment_text}\n</attachment>"

    messages = [
        {
            "role": "system",
            "content": (
                f"{system_prompt}\n\nAfter private reasoning, answer once with only the "
                "requested deliverable. Follow the user's output format exactly. Do not "
                "restate analysis, reasoning, or self-correction in the final answer."
            ),
        },
        {"role": "user", "content": user_content},
    ]
    completion = model.create_chat_completion(
        messages=messages,
        max_tokens=512,
        temperature=0.6,
        top_p=0.95,
        top_k=20,
    )
    return completion["choices"][0]["message"]["content"].rsplit("</think>", 1)[-1].strip()


CSS = """
.gradio-container { max-width: 1180px !important; }
.agent-panel { border: 2px dashed #79b5ce; border-radius: 18px; padding: 8px; }
.output-panel { border: 2px dashed #f0aeb7; border-radius: 18px; padding: 8px; }
"""

with gr.Blocks(css=CSS, title="Spreadsheet Data Agent") as demo:
    gr.Markdown(
        """
        # Spreadsheet Data Agent

        [Code](https://github.com/electblake/Spreadsheet-RL-Data-Agent) | [Demo](https://huggingface.co/spaces/electblake/spreadsheet-data-agent) | [Paper](https://arxiv.org/abs/2605.22642) | [Spreadsheet-RL Model](https://huggingface.co/Spreadsheet-RL/Spreadsheet-RL-4B)

        Send instructions and optional file context to Spreadsheet-RL-4B. This first
        inference surface implements the prompt-and-file entry point from the agent diagram.
        """
    )

    with gr.Row():
        with gr.Column(scale=1, elem_classes="agent-panel"):
            gr.Markdown("### RL data input")
            system_prompt = gr.Textbox(
                label="System prompt",
                value=(
                    "You are a spreadsheet reasoning assistant. Inspect the supplied "
                    "spreadsheet or text context and answer the user's request precisely."
                ),
                lines=5,
            )
            user_prompt = gr.Textbox(
                label="User prompt",
                placeholder="Describe the spreadsheet task or ask a question…",
                lines=8,
            )
            attachment = gr.File(
                label="Optional file context",
                file_types=[".txt", ".md", ".json", ".csv", ".tsv", ".xlsx", ".xls"],
                type="filepath",
            )
            quantization = gr.Dropdown(
                choices=list(QUANT_FILES),
                value="Q4_K_M · 2.8 GB · recommended",
                label="Spreadsheet-RL-4B quantization",
                info="Static GGUF quants published by mradermacher; Q4_K_M is the reference recommendation.",
            )
            run = gr.Button("Run inference", variant="primary")

        with gr.Column(scale=1, elem_classes="output-panel"):
            gr.Markdown("### Agent response")
            response = gr.Textbox(
                label="Generated text",
                lines=28,
                buttons=["copy"],
            )

    gr.Markdown(
        """
        ---

        ### Citation

        If you use Spreadsheet-RL-4B, please cite the model's paper:

        ```bibtex
        @misc{chi2026spreadsheetrl,
          title         = {Spreadsheet-RL: Advancing Large Language Model Agents on Realistic Spreadsheet Tasks via Reinforcement Learning},
          author        = {Banghao Chi and Yining Xie and Mingyuan Wu and Jingcheng Yang and Jize Jiang and Zhaoheng Li and Shengyi Qian and Minjia Zhang and Klara Nahrstedt and Rui Hou and Xiangjun Fan and Hanchao Yu},
          year          = {2026},
          eprint        = {2605.22642},
          archivePrefix = {arXiv},
          primaryClass  = {cs.AI},
          doi           = {10.48550/arXiv.2605.22642},
          url           = {https://arxiv.org/abs/2605.22642}
        }
        ```

        Citation from the [Spreadsheet-RL-4B model card](https://huggingface.co/Spreadsheet-RL/Spreadsheet-RL-4B#citation).
        """
    )

    run.click(
        fn=download_quant,
        inputs=quantization,
        outputs=None,
        show_progress="full",
    ).then(
        fn=generate,
        inputs=[system_prompt, user_prompt, attachment, quantization],
        outputs=response,
        api_name="generate",
        show_progress="full",
    )

demo.queue().launch(mcp_server=True)