Initial Gradio demo for robuchan recipe adapter

README.md

@@ -1,12 +1,12 @@
 ---
-title: Robuchan Demo
-emoji: 📚
-colorFrom: purple
+title: Robuchan - Recipe Adaptation
+emoji: "\U0001F35C"
+colorFrom: yellow
 colorTo: red
 sdk: gradio
-sdk_version: 6.8.0
+sdk_version: "5.33.0"
 app_file: app.py
 pinned: false
+license: apache-2.0
+suggested_hardware: t4-small
 ---
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

RUNBOOK.md

@@ -0,0 +1,136 @@
+# Robuchan HF Space Deployment Runbook
+
+Deploy the Gradio demo to `sumitdotml/robuchan-demo` on Hugging Face Spaces.
+
+## Prerequisites
+
+- `hf` CLI installed ([docs](https://huggingface.co/docs/huggingface_hub/en/guides/cli))
+- Authenticated with a write-access token
+- Files ready in `demo/space/`: `app.py`, `requirements.txt`, `README.md`
+
+## Step 0: Install hf CLI (if needed)
+
+```bash
+curl -LsSf https://hf.co/cli/install.sh | bash
+```
+
+Or via uvx (no install needed):
+
+```bash
+uvx hf --help
+```
+
+## Step 1: Authenticate
+
+```bash
+hf auth login
+# Paste your token when prompted (needs write access)
+# Say yes to saving as git credential
+```
+
+Verify:
+
+```bash
+hf auth whoami
+# Should print: sumitdotml
+```
+
+## Step 2: Create the Space
+
+```bash
+hf repos create robuchan-demo --repo-type space --space-sdk gradio
+```
+
+Expected output:
+
+```
+Successfully created sumitdotml/robuchan-demo on the Hub.
+Your repo is now available at https://huggingface.co/spaces/sumitdotml/robuchan-demo
+```
+
+If the Space already exists, add `--exist-ok`:
+
+```bash
+hf repos create robuchan-demo --repo-type space --space-sdk gradio --exist-ok
+```
+
+## Step 3: Upload files
+
+From the repo root:
+
+```bash
+hf upload sumitdotml/robuchan-demo demo/space . --repo-type space \
+  --commit-message "Initial Gradio demo for robuchan recipe adapter"
+```
+
+This uploads the contents of `demo/space/` (app.py, requirements.txt, README.md) to the root of the Space repo.
+
+Expected output:
+
+```
+https://huggingface.co/spaces/sumitdotml/robuchan-demo/tree/main/
+```
+
+## Step 4: Set hardware to T4
+
+The `README.md` frontmatter includes `suggested_hardware: t4-small`, but you may need to set it manually in Space settings:
+
+1. Go to https://huggingface.co/spaces/sumitdotml/robuchan-demo/settings
+2. Under **Space Hardware**, select **T4 small**
+3. Click **Save**
+
+(Requires HF Pro or a hardware grant.)
+
+## Step 5: Wait for build
+
+The Space will auto-build on push. Monitor the build log:
+
+1. Go to https://huggingface.co/spaces/sumitdotml/robuchan-demo
+2. Click the **Logs** tab (or the "Building" badge)
+3. Wait for "Running on local URL" in the logs
+
+Build typically takes 3-5 minutes (dependency install + model download on first boot).
+
+## Step 6: Verify
+
+### Quick smoke test
+
+1. Open https://huggingface.co/spaces/sumitdotml/robuchan-demo
+2. Click the first example (tonkotsu ramen, vegan) and hit **Submit**
+3. Wait for generation (~30-60s on T4)
+4. Confirm output contains all 5 sections:
+   - Substitution Plan
+   - Adapted Ingredients
+   - Adapted Steps
+   - Flavor Preservation Notes
+   - Constraint Check
+
+### Full verification
+
+Run both pre-loaded examples:
+
+| Example | Constraint | Check |
+|---------|-----------|-------|
+| Tonkotsu ramen | vegan | No pork/eggs/animal products in adapted recipe |
+| Japanese curry | gluten_free | No wheat flour/soy sauce in adapted recipe |
+
+Also test a custom input: paste any recipe, select a constraint, verify structured output.
+
+## Updating the Space
+
+After editing files in `demo/space/`, re-upload:
+
+```bash
+hf upload sumitdotml/robuchan-demo demo/space . --repo-type space \
+  --commit-message "description of changes"
+```
+
+## Troubleshooting
+
+| Symptom | Fix |
+|---------|-----|
+| Build fails on `bitsandbytes` | Needs CUDA runtime. Verify hardware is set to T4, not CPU. |
+| OOM during model load | T4 has 16GB VRAM. 4-bit quantization should fit 8B model. If OOM, check no other process is using GPU. Restart Space. |
+| "Model not found" error | Verify `sumitdotml/robuchan` adapter is public (or set `HF_TOKEN` as a Space secret). |
+| Space stuck on "Building" | Check build logs for pip install errors. May need to pin a specific torch version in requirements.txt. |
+| Slow first inference | Expected. First request triggers CUDA kernel compilation. Subsequent requests are faster. |

app.py

@@ -0,0 +1,169 @@
+"""Robuchan - Recipe Adaptation Demo (HF Space).
+
+Loads the fine-tuned LoRA adapter with 4-bit quantization and serves
+a Gradio interface for interactive recipe adaptation.
+"""
+
+from __future__ import annotations
+
+import gradio as gr
+import torch
+from peft import AutoPeftModelForCausalLM
+from transformers import AutoTokenizer, BitsAndBytesConfig
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+ADAPTER_ID = "sumitdotml/robuchan"
+BASE_MODEL = "mistralai/Ministral-8B-Instruct-2410"
+MAX_NEW_TOKENS = 800
+
+SYSTEM_PROMPT = (
+    "You are a culinary adaptation assistant. Priority: (1) strict dietary "
+    "compliance, (2) preserve dish identity and flavor profile, (3) keep "
+    "instructions practical and cookable. Never include forbidden ingredients "
+    "or their derivatives (stocks, sauces, pastes, broths). If no exact "
+    "compliant substitute exists, acknowledge the gap, choose the closest "
+    "viable option, and state the trade-off. Output sections exactly: "
+    "Substitution Plan, Adapted Ingredients, Adapted Steps, Flavor "
+    "Preservation Notes, Constraint Check."
+)
+
+CONSTRAINTS = [
+    "vegan",
+    "dairy_free",
+    "gluten_free",
+    "vegetarian",
+    "nut_free",
+    "egg_free",
+    "low_sodium",
+]
+
+EXAMPLE_TONKOTSU = (
+    "Adapt this tonkotsu ramen recipe for a vegan diet:\n\n"
+    "Ingredients: pork bones, pork belly, eggs, wheat noodles, soy sauce, "
+    "mirin, garlic, ginger, green onions, nori, sesame oil.\n\n"
+    "Instructions: Boil pork bones for 12 hours for broth. Char pork belly. "
+    "Soft-boil eggs. Cook wheat noodles. Assemble with toppings."
+)
+
+EXAMPLE_CURRY = (
+    "Adapt this Japanese curry recipe for gluten-free:\n\n"
+    "Ingredients: curry roux (contains wheat flour), chicken thighs, potatoes, "
+    "carrots, onions, rice, soy sauce, mirin, dashi stock.\n\n"
+    "Instructions: Saut\u00e9 onions, brown chicken, add vegetables, add water and "
+    "curry roux, simmer 20 minutes. Serve over rice."
+)
+
+# ---------------------------------------------------------------------------
+# Model loading (once at startup)
+# ---------------------------------------------------------------------------
+
+print("Loading model with 4-bit quantization ...")
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_use_double_quant=True,
+)
+
+model = AutoPeftModelForCausalLM.from_pretrained(
+    ADAPTER_ID,
+    quantization_config=bnb_config,
+    device_map="auto",
+    torch_dtype=torch.bfloat16,
+)
+model.eval()
+
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+print("Model loaded.")
+
+# ---------------------------------------------------------------------------
+# Generation
+# ---------------------------------------------------------------------------
+
+
+def generate(recipe: str, constraint: str) -> str:
+    """Generate an adapted recipe from the fine-tuned model."""
+    if not recipe.strip():
+        return "Please paste a recipe above."
+
+    user_content = f"Constraint: {constraint}\n\n{recipe}"
+    messages = [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "user", "content": user_content},
+    ]
+
+    text = tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True,
+    )
+    inputs = tokenizer(text, return_tensors="pt")
+    inputs = {k: v.to(model.device) for k, v in inputs.items()}
+
+    gen_kwargs = {
+        "max_new_tokens": MAX_NEW_TOKENS,
+        "do_sample": False,
+        "pad_token_id": tokenizer.pad_token_id,
+    }
+    if tokenizer.eos_token_id is not None:
+        gen_kwargs["eos_token_id"] = tokenizer.eos_token_id
+
+    with torch.no_grad():
+        output_ids = model.generate(**inputs, **gen_kwargs)
+
+    prompt_len = inputs["input_ids"].shape[1]
+    completion_ids = output_ids[0][prompt_len:]
+    return tokenizer.decode(completion_ids, skip_special_tokens=True).strip()
+
+
+# ---------------------------------------------------------------------------
+# Gradio UI
+# ---------------------------------------------------------------------------
+
+DESCRIPTION = """\
+# Robuchan - Recipe Adaptation
+
+Fine-tuned [Ministral 8B](https://huggingface.co/mistralai/Ministral-8B-Instruct-2410) \
+adapter for dietary-compliant recipe transformation.
+
+**How it works:** paste a recipe, pick a dietary constraint, and the model \
+generates an adapted version with substitution rationale, modified ingredients, \
+updated steps, and a compliance check.
+
+Adapter: [`sumitdotml/robuchan`](https://huggingface.co/sumitdotml/robuchan)
+"""
+
+examples = [
+    [EXAMPLE_TONKOTSU, "vegan"],
+    [EXAMPLE_CURRY, "gluten_free"],
+]
+
+demo = gr.Interface(
+    fn=generate,
+    inputs=[
+        gr.Textbox(
+            label="Recipe",
+            placeholder="Paste ingredients + instructions here ...",
+            lines=10,
+        ),
+        gr.Dropdown(
+            choices=CONSTRAINTS,
+            value="vegan",
+            label="Dietary Constraint",
+        ),
+    ],
+    outputs=gr.Markdown(label="Adapted Recipe"),
+    title="Robuchan",
+    description=DESCRIPTION,
+    examples=examples,
+    cache_examples=False,
+    flagging_mode="never",
+)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,7 @@
+torch
+transformers
+peft
+accelerate
+bitsandbytes
+gradio
+huggingface-hub