Switch to SmolLM2-360M (prebuilt wheels, no C++ compile)

README.md

@@ -14,8 +14,9 @@ tags:
   - tiny-titan
   - best-agent
   - off-brand
+  - openai-codex
 models:
-  - hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF
+  - HuggingFaceTB/SmolLM2-360M-Instruct
 ---
 
 # 🫀 Body Debt
@@ -33,16 +34,16 @@ Body Debt calculates the precise recovery cost of last night's choices — alcoh
 
 ## The model
 
-**Llama-3.2-1B-Instruct** (Q4_K_M quantization, ~700MB) — runs entirely on CPU via `llama-cpp-python`. No API calls, no cloud inference. Your health data never leaves your machine.
+**SmolLM2-360M-Instruct** (360M parameters) — runs entirely on CPU via HuggingFace Transformers. No external API calls, no cloud inference. Your health data stays on-device.
 
 The face scan stress classifier is a custom 7→16→8→1 MLP (~2KB ONNX) that converts facial geometry features into a fatigue score.
 
 ## Tech
 
-- **LLM**: Llama-3.2-1B-Instruct (1B params, Q4_K_M GGUF) via llama-cpp-python
+- **LLM**: SmolLM2-360M-Instruct (360M params) via HuggingFace Transformers
 - **Face analysis**: MediaPipe FaceMesh → 7 stress features → ONNX MLP
 - **Scoring**: Deterministic 5-system engine with physiological weights, drink-type modifiers, training CNS load, circadian alignment penalties
-- **UI**: Gradio 5 with custom dark theme
+- **UI**: Gradio 6 with custom dark theme
 
 ## Privacy
 
@@ -66,9 +67,15 @@ python generate_model.py  # creates the ONNX stress model
 python app.py
 ```
 
+## OpenAI Codex Track
+
+This Space was built with OpenAI Codex as the coding agent. The public source repository, including Codex-attributed commits, is here:
+
+**Repository:** [github.com/udirobert/bodydebt](https://github.com/udirobert/bodydebt)
+
 ## Full product
 
-The complete Body Debt application (Next.js, ZK proofs on SKALE, real-time animated dashboard) is at: [github.com/body-debt](https://github.com)
+The complete Body Debt application (Next.js, ZK proofs on SKALE, real-time animated dashboard) is at: [github.com/udirobert/bodydebt](https://github.com/udirobert/bodydebt)
 
 ---
 

app.py

@@ -198,7 +198,7 @@ def run_analysis(
     progress(1.0, desc="Done!")
 
     advice_md = "### 🤖 Recovery Prescription\n\n"
-    advice_md += f"*Generated by Llama-3.2-1B running locally*\n\n{advice}"
+    advice_md += f"*Generated by SmolLM2-360M running locally*\n\n{advice}"
 
     return score_md, face_text, advice_md
 
@@ -323,7 +323,7 @@ with gr.Blocks(title="Body Debt") as demo:
     gr.Markdown(
         """
         ---
-        *Body Debt uses Llama-3.2-1B (1 billion parameters) running locally via llama-cpp-python.
+        *Body Debt uses SmolLM2-360M-Instruct (360M parameters) running locally via HuggingFace Transformers.
         Face analysis uses MediaPipe FaceMesh — no biometric data leaves your device.
         Built for the [Build Small Hackathon](https://huggingface.co/spaces/huggingface/build-small-hackathon).*
         """

health_coach.py

@@ -1,5 +1,5 @@
 """
-Local LLM health coach using llama-cpp-python.
+Local LLM health coach using HuggingFace Transformers.
 Generates personalized recovery advice from stressor + face scan data.
 Falls back to a template-based response if model unavailable.
 """
@@ -7,27 +7,9 @@ Falls back to a template-based response if model unavailable.
 from __future__ import annotations
 
 import os
-from pathlib import Path
 from typing import Optional
 
-from huggingface_hub import hf_hub_download
-
-MODEL_REPO = "hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF"
-MODEL_FILE = "llama-3.2-1b-instruct-q4_k_m.gguf"
-CACHE_DIR = Path.home() / ".cache" / "body-debt-models"
-
-
-def get_model_path() -> Path:
-    local = CACHE_DIR / MODEL_FILE
-    if local.exists():
-        return local
-    CACHE_DIR.mkdir(parents=True, exist_ok=True)
-    path = hf_hub_download(
-        repo_id=MODEL_REPO,
-        filename=MODEL_FILE,
-        local_dir=str(CACHE_DIR),
-    )
-    return Path(path)
+MODEL_ID = "HuggingFaceTB/SmolLM2-360M-Instruct"
 
 
 def generate_advice(
@@ -37,66 +19,70 @@ def generate_advice(
     face_stress: Optional[float] = None,
     progress_callback=None,
 ) -> str:
-    """Generate personalized recovery advice using local Llama-3.2-1B."""
+    """Generate personalized recovery advice using a small local LLM."""
     try:
         if progress_callback:
             progress_callback(0.1, "Loading model...")
-        model_path = get_model_path()
-        if progress_callback:
-            progress_callback(0.5, "Model loaded, generating advice...")
-        return _llm_generate(model_path, debt_score, system_scores, stressor_summary, face_stress)
+        return _transformers_generate(debt_score, system_scores, stressor_summary, face_stress, progress_callback)
     except Exception as e:
+        print(f"LLM generation failed: {e}")
         return _fallback_advice(debt_score, system_scores, stressor_summary)
 
 
-def _build_prompt(
+def _build_messages(
     debt_score: int,
     system_scores: list[dict],
     stressor_summary: str,
     face_stress: Optional[float],
-) -> str:
+) -> list[dict]:
     systems_text = "\n".join(
         f"- {s['label']}: {s['score']}/100 (clears {s['cleared_at']})"
         for s in system_scores
     )
     face_text = f"\nFacial stress indicator: {face_stress:.0f}/100" if face_stress else ""
 
-    return f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>
-You are a concise recovery coach. Given physiological debt data, provide specific, actionable recovery advice in 4 categories: Right Now, This Morning, Today, Avoid. Be direct, no fluff. Use the system scores to prioritize which body systems need attention most urgently.<|eot_id|><|start_header_id|>user<|end_header_id|>
-My body debt score: {debt_score}/100
-Stressors: {stressor_summary}{face_text}
-
-System breakdown:
-{systems_text}
-
-Give me my recovery prescription.<|eot_id|><|start_header_id|>assistant<|end_header_id|>
-"""
+    return [
+        {
+            "role": "system",
+            "content": "You are a concise recovery coach. Given physiological debt data, provide specific, actionable recovery advice in 4 categories: Right Now, This Morning, Today, Avoid. Be direct, no fluff. Use the system scores to prioritize which body systems need attention most urgently.",
+        },
+        {
+            "role": "user",
+            "content": f"My body debt score: {debt_score}/100\nStressors: {stressor_summary}{face_text}\n\nSystem breakdown:\n{systems_text}\n\nGive me my recovery prescription.",
+        },
+    ]
 
 
-def _llm_generate(
-    model_path: Path,
+def _transformers_generate(
     debt_score: int,
     system_scores: list[dict],
     stressor_summary: str,
     face_stress: Optional[float],
+    progress_callback=None,
 ) -> str:
-    from llama_cpp import Llama
+    from transformers import pipeline
 
-    llm = Llama(
-        model_path=str(model_path),
-        n_ctx=2048,
-        n_threads=4,
-        verbose=False,
-    )
-    prompt = _build_prompt(debt_score, system_scores, stressor_summary, face_stress)
-    output = llm(
-        prompt,
-        max_tokens=512,
-        temperature=0.7,
-        top_p=0.9,
-        stop=["<|eot_id|>"],
+    if progress_callback:
+        progress_callback(0.3, "Loading SmolLM2-360M...")
+
+    pipe = pipeline(
+        "text-generation",
+        model=MODEL_ID,
+        device_map="auto",
+        torch_dtype="auto",
     )
-    return output["choices"][0]["text"].strip()
+
+    if progress_callback:
+        progress_callback(0.6, "Generating advice...")
+
+    messages = _build_messages(debt_score, system_scores, stressor_summary, face_stress)
+    output = pipe(messages, max_new_tokens=300, temperature=0.7, do_sample=True)
+    result = output[0]["generated_text"][-1]["content"]
+
+    if progress_callback:
+        progress_callback(1.0, "Done!")
+
+    return result
 
 
 def _fallback_advice(debt_score: int, system_scores: list[dict], stressor_summary: str) -> str:

requirements.txt

@@ -2,5 +2,6 @@ gradio>=5.0.0,<7.0.0
 mediapipe>=0.10.14
 numpy>=1.26.0
 onnxruntime>=1.18.0
-llama-cpp-python>=0.3.0
 huggingface_hub>=0.25.0
+transformers>=4.45.0
+accelerate>=1.0.0