one-shot 5.6 sol

.gitignore

@@ -0,0 +1,11 @@
+# Python-generated files
+__pycache__/
+*.py[oc]
+build/
+dist/
+wheels/
+*.egg-info
+
+# Virtual environments
+.venv
+uv.lock

README.md

@@ -0,0 +1,22 @@
+---
+title: Spreadsheet Data Agent
+emoji: 📊
+colorFrom: blue
+colorTo: pink
+sdk: gradio
+sdk_version: 6.24.0
+app_file: app.py
+pinned: false
+license: apache-2.0
+models:
+  - Spreadsheet-RL/Spreadsheet-RL-4B
+  - mradermacher/Spreadsheet-RL-4B-GGUF
+---
+
+# Spreadsheet Data Agent
+
+A basic text-and-file inference app for Spreadsheet-RL-4B, modeled on the prompt entry point in the Spreadsheet-RL agent-system diagram.
+
+The app accepts a system prompt, user prompt, and optional text or spreadsheet file. Its quantization selector exposes the 4B GGUF variants captured in the project reference material, with Q4_K_M selected by default.
+
+ZeroGPU support is enabled with the `spaces` package and `@spaces.GPU`. Select ZeroGPU in the Hugging Face Space hardware settings after deployment.

app.py

@@ -0,0 +1,185 @@
+from gc import collect
+from pathlib import Path
+
+import gradio as gr
+import pandas as pd
+import spaces
+import torch
+from huggingface_hub import hf_hub_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+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
+tokenizer = 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, tokenizer
+
+    quant_file = QUANT_FILES[quantization]
+    if active_quant != quantization:
+        model = None
+        tokenizer = None
+        active_quant = None
+        collect()
+        torch.cuda.empty_cache()
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            MODEL_REPO,
+            gguf_file=quant_file,
+        )
+        model = AutoModelForCausalLM.from_pretrained(
+            MODEL_REPO,
+            gguf_file=quant_file,
+            dtype=torch.bfloat16,
+            device_map="cuda",
+        )
+        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": system_prompt},
+        {"role": "user", "content": user_content},
+    ]
+    inputs = tokenizer.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        return_tensors="pt",
+    ).to(model.device)
+
+    with torch.inference_mode():
+        generated = model.generate(
+            inputs,
+            max_new_tokens=512,
+            do_sample=True,
+            temperature=0.6,
+            top_p=0.95,
+            top_k=20,
+        )
+
+    return tokenizer.decode(
+        generated[0, inputs.shape[-1] :],
+        skip_special_tokens=True,
+    )
+
+
+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
+        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"],
+            )
+
+    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()

mise.toml

@@ -0,0 +1,7 @@
+[tools]
+powershell = "7"
+python = "3.12"
+uv = "latest"
+
+[env]
+UV_LINK_MODE="copy"

pyproject.toml

@@ -0,0 +1,25 @@
+[project]
+name = "spreadsheetapp"
+version = "0.1.0"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.12.10"
+dependencies = [
+    "accelerate>=1.14.0",
+    "gguf>=0.19.0",
+    "huggingface-hub>=1.27.0",
+    "openpyxl>=3.1.5",
+    "pandas>=3.0.5",
+    "spaces>=0.51.1",
+    "torch==2.11.0",
+    "transformers==5.15.0",
+    "xlrd>=2.0.2",
+]
+
+[tool.uv.sources]
+torch = { index = "pytorch-cu130" }
+
+[[tool.uv.index]]
+name = "pytorch-cu130"
+url = "https://download.pytorch.org/whl/cu130"
+explicit = true

requirements.txt

@@ -0,0 +1,9 @@
+accelerate
+gguf
+huggingface-hub
+openpyxl
+pandas
+spaces
+torch==2.11.0
+transformers==5.15.0
+xlrd