Upload folder using huggingface_hub

Dockerfile

@@ -0,0 +1,17 @@
+FROM python:3.11-slim
+
+RUN apt-get update && apt-get install -y build-essential cmake && rm -rf /var/lib/apt/lists/*
+RUN pip install --no-cache-dir \
+    "gradio==4.44.1" \
+    "huggingface-hub==0.26.5" \
+    "llama-cpp-python>=0.3.0"
+
+RUN useradd -m -u 1000 user
+RUN mkdir -p /tmp/hf_cache && chown -R user:user /tmp/hf_cache
+USER user
+WORKDIR /app
+COPY app.py aeon.css ./
+ENV PYTHONUNBUFFERED=1
+ENV HF_HOME=/tmp/hf_cache
+EXPOSE 7860
+CMD ["python", "app.py"]

README.md

@@ -1,15 +1,9 @@
 ---
 title: The Void - Buleyean RL Demo
-emoji: "\U0001F573\uFE0F"
+emoji: 🕳️
 colorFrom: gray
 colorTo: indigo
-sdk: gradio
-sdk_version: 5.23.0
-python_version: "3.11"
-app_file: app.py
+sdk: docker
+app_port: 7860
 pinned: true
-models:
-  - bartowski/SmolLM2-360M-Instruct-GGUF
-  - forkjoin-ai/buleyean-smollm2-360m
-  - HuggingFaceTB/SmolLM2-360M-Instruct
 ---

aeon.css

@@ -0,0 +1,148 @@
+/* AeonOS Design System for Gradio Spaces */
+
+.gradio-container {
+  background: #09090b !important;
+  color: #fafafa !important;
+  font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif !important;
+}
+
+/* Surface panels */
+.block, .panel, .form, .wrap, .contain {
+  background: #0c0c0f !important;
+  border-color: #1f1f23 !important;
+  border-radius: 6px !important;
+}
+
+/* Elevated panels */
+.gr-panel, .gr-box, .gr-form {
+  background: #111114 !important;
+  border: 1px solid #1f1f23 !important;
+}
+
+/* Text */
+.prose, .markdown, label, .label-wrap {
+  color: #fafafa !important;
+}
+
+.prose h1, .prose h2, .prose h3, .markdown h1, .markdown h2, .markdown h3 {
+  color: #fafafa !important;
+}
+
+.prose p, .markdown p {
+  color: #a1a1aa !important;
+}
+
+/* Secondary text */
+.secondary-text, .caption, .info {
+  color: #71717a !important;
+}
+
+/* Links */
+a, .prose a, .markdown a {
+  color: #3b82f6 !important;
+  text-decoration: none !important;
+}
+a:hover {
+  color: #60a5fa !important;
+}
+
+/* Inputs */
+input, textarea, select, .gr-input, .gr-text-input {
+  background: #09090b !important;
+  border: 1px solid #1f1f23 !important;
+  color: #fafafa !important;
+  border-radius: 6px !important;
+  font-size: 16px !important;
+}
+input:focus, textarea:focus {
+  border-color: #3b82f6 !important;
+  outline: none !important;
+  box-shadow: 0 0 0 2px rgba(59, 130, 246, 0.15) !important;
+}
+
+/* Primary button */
+.primary, button.primary, .gr-button-primary {
+  background: #3b82f6 !important;
+  color: #fafafa !important;
+  border: none !important;
+  border-radius: 6px !important;
+  font-weight: 500 !important;
+  transition: background 150ms ease !important;
+}
+.primary:hover, button.primary:hover {
+  background: #60a5fa !important;
+}
+
+/* Secondary buttons */
+button, .gr-button {
+  background: #18181b !important;
+  color: #a1a1aa !important;
+  border: 1px solid #1f1f23 !important;
+  border-radius: 6px !important;
+}
+button:hover, .gr-button:hover {
+  background: #1e1e22 !important;
+  border-color: #27272a !important;
+}
+
+/* Tables */
+table {
+  border-collapse: collapse !important;
+}
+th {
+  background: #111114 !important;
+  color: #a1a1aa !important;
+  border-bottom: 1px solid #1f1f23 !important;
+  padding: 8px 12px !important;
+  text-align: left !important;
+}
+td {
+  border-bottom: 1px solid #17171a !important;
+  padding: 8px 12px !important;
+  color: #fafafa !important;
+}
+
+/* Code blocks */
+code, pre {
+  background: #111114 !important;
+  color: #93c5fd !important;
+  border-radius: 4px !important;
+  border: 1px solid #1f1f23 !important;
+}
+
+/* Accent glow for key metrics */
+.metric-highlight {
+  color: #3b82f6 !important;
+  font-weight: 600 !important;
+}
+
+/* Status colors */
+.success { color: #22c55e !important; }
+.warning { color: #eab308 !important; }
+.error { color: #ef4444 !important; }
+
+/* Dataset/Examples component */
+.dataset-row, .sample-row {
+  background: #111114 !important;
+  border: 1px solid #1f1f23 !important;
+}
+.dataset-row:hover, .sample-row:hover {
+  background: #18181b !important;
+}
+
+/* Scrollbar */
+::-webkit-scrollbar {
+  width: 6px;
+}
+::-webkit-scrollbar-track {
+  background: #09090b;
+}
+::-webkit-scrollbar-thumb {
+  background: #27272a;
+  border-radius: 3px;
+}
+
+/* Footer */
+footer {
+  color: #52525b !important;
+}

app.py

@@ -4,219 +4,84 @@ LIVE inference only. Every response generated in real-time.
 """
 
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftModel
-import torch
-
-BASE_MODEL_ID = "HuggingFaceTB/SmolLM2-360M-Instruct"
-BULEYEAN_ADAPTER = "forkjoin-ai/buleyean-smollm2-360m"
-
-print("Loading tokenizer...", flush=True)
-tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
-if tokenizer.pad_token is None:
-    tokenizer.pad_token = tokenizer.eos_token
-
-print(f"Loading base model ({BASE_MODEL_ID})...", flush=True)
-base_model = AutoModelForCausalLM.from_pretrained(
-    BASE_MODEL_ID,
-    torch_dtype=torch.float32,
-    device_map="cpu",
-    trust_remote_code=True,
+from llama_cpp import Llama
+from huggingface_hub import hf_hub_download
+import os, sys
+
+print("Downloading Buleyean model...", flush=True)
+bule_path = hf_hub_download(
+    repo_id="forkjoin-ai/buleyean-smollm2-360m",
+    filename="buleyean-smollm2-360m-q4_k_m.gguf",
+    cache_dir="/tmp/hf_cache",
 )
-print("Base model loaded.", flush=True)
-
-print("Loading Buleyean adapter...", flush=True)
-bule_base = AutoModelForCausalLM.from_pretrained(
-    BASE_MODEL_ID,
-    torch_dtype=torch.float32,
-    device_map="cpu",
-    trust_remote_code=True,
+print(f"Buleyean model ready.", flush=True)
+
+print("Downloading base model...", flush=True)
+base_path = hf_hub_download(
+    repo_id="bartowski/SmolLM2-360M-Instruct-GGUF",
+    filename="SmolLM2-360M-Instruct-Q4_K_M.gguf",
+    cache_dir="/tmp/hf_cache",
 )
-try:
-    bule_model = PeftModel.from_pretrained(bule_base, BULEYEAN_ADAPTER)
-    bule_model = bule_model.merge_and_unload()
-    print("Buleyean adapter merged.", flush=True)
-except Exception as e:
-    print(f"Warning: Could not load adapter ({e}), using base model copy", flush=True)
-    bule_model = bule_base
+print(f"Base model ready.", flush=True)
 
-print("All models ready. Live inference active.", flush=True)
+print("Loading models into memory...", flush=True)
+bule_llm = Llama(model_path=bule_path, n_ctx=512, n_threads=4, verbose=False)
+base_llm = Llama(model_path=base_path, n_ctx=512, n_threads=4, verbose=False)
+print("Both models loaded. Live inference ready.", flush=True)
 
 
-def generate(prompt, model, max_tokens=300, temperature=0.7, top_p=0.9):
-    messages = [{"role": "user", "content": prompt}]
-    input_text = tokenizer.apply_chat_template(
-        messages, tokenize=False, add_generation_prompt=True
-    )
-    inputs = tokenizer(input_text, return_tensors="pt")
-    with torch.no_grad():
-        outputs = model.generate(
-            **inputs,
-            max_new_tokens=max_tokens,
-            temperature=temperature,
-            top_p=top_p,
-            do_sample=True,
-            pad_token_id=tokenizer.pad_token_id,
-        )
-    response = tokenizer.decode(
-        outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True
+def generate(prompt, model):
+    out = model(
+        f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n",
+        max_tokens=300,
+        temperature=0.7,
+        top_p=0.9,
+        stop=["<|im_end|>", "<|im_start|>"],
     )
-    return response.strip()
+    return out["choices"][0]["text"].strip()
 
 
 def compare(prompt):
     if not prompt or not prompt.strip():
         return "Please enter a prompt.", "Please enter a prompt."
-    base_out = generate(prompt, base_model)
-    bule_out = generate(prompt, bule_model)
+    base_out = generate(prompt, base_llm)
+    bule_out = generate(prompt, bule_llm)
     return base_out, bule_out
 
 
-# ---------------------------------------------------------------------------
-# UI
-# ---------------------------------------------------------------------------
-
-with gr.Blocks(
-    title="The Void -- Buleyean RL",
-    theme=gr.themes.Base(primary_hue="indigo"),
-) as demo:
-
+import pathlib; _aeon_css = pathlib.Path("aeon.css").read_text() if pathlib.Path("aeon.css").exists() else ""
+with gr.Blocks(css=_aeon_css, title="The Void", theme=gr.themes.Base(primary_hue="indigo")) as demo:
     gr.Markdown("""# The Void -- Buleyean RL
 
 **Live inference. Every response generated in real-time. Nothing hardcoded. Nothing fabricated.**
 
-Standard RLHF learns what to say by imitating chosen completions.
-Buleyean RL learns what *not* to say by studying rejections.
-The complement distribution preserves the (K-1) rejected perspectives.
-
-[Library](https://github.com/forkjoin-ai/buleyean-rl) | [Paper](https://forkracefold.com) | [Training Data](https://huggingface.co/datasets/forkjoin-ai/buleyean-rejection-data) | 500+ Lean 4 theorems, zero sorry
-""")
+Type any prompt. Both models run inference right now on this machine.
 
-    with gr.Tab("Compare (Base vs Buleyean)"):
-        gr.Markdown("""Type any prompt. Both models run inference right now on this machine.
+Base: [SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct) (Q4_K_M GGUF)
+Buleyean: [buleyean-smollm2-360m](https://huggingface.co/forkjoin-ai/buleyean-smollm2-360m) -- same model, trained from rejection alone (Q4_K_M GGUF)
 
-**Base:** [SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct) |
-**Buleyean:** [buleyean-smollm2-360m](https://huggingface.co/forkjoin-ai/buleyean-smollm2-360m) -- same architecture, trained from rejection alone
-""")
-        prompt = gr.Textbox(
-            label="Your prompt", lines=2,
-            placeholder="Type anything and press Generate...",
-        )
-        btn = gr.Button("Generate (live inference)", variant="primary", size="lg")
-        with gr.Row():
-            with gr.Column():
-                gr.Markdown("### Base Model (null hypothesis)")
-                base_out = gr.Textbox(
-                    label="SmolLM2-360M-Instruct", lines=12, interactive=False,
-                )
-            with gr.Column():
-                gr.Markdown("### Buleyean-Trained (from the void)")
-                bule_out = gr.Textbox(
-                    label="buleyean-smollm2-360m", lines=12, interactive=False,
-                )
-
-        btn.click(compare, [prompt], [base_out, bule_out])
-        prompt.submit(compare, [prompt], [base_out, bule_out])
-
-        gr.Markdown("### Try these prompts:")
-        for p in [
-            "What is the theory of failure?",
-            "How are you feeling today?",
-            "I've been feeling really anxious lately.",
-            "Write me a haiku about failure.",
-            "What is the meaning of life?",
-        ]:
-            gr.Button(p, size="sm").click(
-                fn=lambda x=p: compare(x),
-                inputs=[],
-                outputs=[base_out, bule_out],
-            ).then(fn=lambda x=p: x, inputs=[], outputs=[prompt])
-
-    with gr.Tab("Personality Models"):
-        gr.Markdown("""## The Personality IS the Walker
-
-Same rejection data. Same base model. Five different complement distributions.
-
-Each personality profile modulates how the void walker traverses the rejection boundary:
-
-| Personality | Try (Fork) | Choose (Race) | Commit (Fold) | Let Go (Vent) | Learn (Interfere) | Result |
-|---|---|---|---|---|---|---|
-| **Builder** | 0.5 | 0.8 | **0.9** | 0.4 | 0.618 | Tightest convergence (97%). The fold dominates. |
-| **Anxious** | 0.3 | 0.5 | 0.7 | **0.15** | 0.4 | Learns slowly, forgets nothing (79%). |
-| **Balanced** | 0.618 | 0.618 | 0.618 | 0.618 | 0.618 | All phi. The control (81%). |
-| **Explorer** | **0.9** | 0.618 | 0.4 | 0.7 | 0.85 | Wide aperture, broad distribution (73%). |
-| **Creative** | **0.95** | 0.4 | 0.3 | 0.8 | **0.9** | Max divergence, keeps options open (73%). |
-
-### Training Results (Qwen2.5-32B-Instruct)
-
-| Personality | Alpha | Final Loss | Min Loss | Curriculum |
-|---|---|---|---|---|
-| Builder | 0.950 | 0.293 | 0.270 | inverse_bule |
-| Anxious | 0.793 | 0.543 | 0.495 | rejection_density |
-| Balanced | 0.700 | 0.830 | 0.741 | rejection_density |
-| Explorer | 0.453 | 2.937 | 2.708 | kurtosis |
-| Creative | 0.340 | 3.525 | 3.239 | kurtosis |
-
-### Download the models
-
-All five personality-modulated LoRA adapters for Qwen2.5-32B:
-
-- [buleyean-qwen2.5-32b-builder](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-builder)
-- [buleyean-qwen2.5-32b-anxious](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-anxious)
-- [buleyean-qwen2.5-32b-balanced](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-balanced)
-- [buleyean-qwen2.5-32b-explorer](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-explorer)
-- [buleyean-qwen2.5-32b-creative](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-creative)
-
-```python
-from transformers import AutoModelForCausalLM
-from peft import PeftModel
-
-base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-32B-Instruct", device_map="auto")
-# Choose your personality:
-model = PeftModel.from_pretrained(base, "forkjoin-ai/buleyean-qwen2.5-32b-builder", subfolder="lora")
-model = model.merge_and_unload()
-```
+[Library](https://github.com/forkjoin-ai/buleyean-rl) | [Paper](https://huggingface.co/forkjoin-ai) | 500+ Lean 4 theorems, zero sorry
 """)
-
-    with gr.Tab("All Models"):
-        gr.Markdown("""## Buleyean RL Model Family
-
-All models trained from rejection alone. No reward model. No chosen examples.
-
-### Base Buleyean Models
-
-| Model | Base | Size | HF Link |
-|---|---|---|---|
-| buleyean-qwen2.5-32b | Qwen2.5-32B-Instruct | 32B | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b) |
-| buleyean-qwen2.5-7b | Qwen2.5-7B-Instruct | 7B | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-7b) |
-| buleyean-deepseek-r1-7b | DeepSeek-R1-Distill-Qwen-7B | 7B | [Download](https://huggingface.co/forkjoin-ai/buleyean-deepseek-r1-7b) |
-| buleyean-mistral-7b | Mistral-7B-Instruct-v0.3 | 7B | [Download](https://huggingface.co/forkjoin-ai/buleyean-mistral-7b) |
-| buleyean-qwen2.5-0.5b | Qwen2.5-0.5B-Instruct | 0.5B | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-0.5b) |
-| buleyean-smollm2-360m | SmolLM2-360M-Instruct | 360M | [Download](https://huggingface.co/forkjoin-ai/buleyean-smollm2-360m) |
-
-### Personality-Modulated (32B)
-
-| Personality | Commit | Final Loss | HF Link |
-|---|---|---|---|
-| Builder | 0.9 | 0.293 | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-builder) |
-| Anxious | 0.7 | 0.543 | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-anxious) |
-| Balanced | 0.618 | 0.830 | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-balanced) |
-| Explorer | 0.4 | 2.937 | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-explorer) |
-| Creative | 0.3 | 3.525 | [Download](https://huggingface.co/forkjoin-ai/buleyean-qwen2.5-32b-creative) |
-
-### Links
-
-- [Training library](https://github.com/forkjoin-ai/buleyean-rl)
-- [Training data](https://huggingface.co/datasets/forkjoin-ai/buleyean-rejection-data)
-- [Paper](https://forkracefold.com)
-- [Colab: Train your own](https://huggingface.co/datasets/forkjoin-ai/buleyean-rejection-data/blob/main/train_70b_colab.ipynb)
-- [Colab: Personality sweep](https://huggingface.co/datasets/forkjoin-ai/buleyean-rejection-data/blob/main/personality_sweep_colab.ipynb)
-""")
-
-    gr.Markdown("""---
-*Built with [Buleyean RL](https://github.com/forkjoin-ai/buleyean-rl). The void is where the information is.*
-""")
-
+    prompt = gr.Textbox(label="Your prompt", lines=2, placeholder="Type anything and press Generate...")
+    btn = gr.Button("Generate (live inference)", variant="primary", size="lg")
+    with gr.Row():
+        with gr.Column():
+            gr.Markdown("### Base Model (null hypothesis)")
+            base_out = gr.Textbox(label="SmolLM2-360M-Instruct", lines=12, interactive=False)
+        with gr.Column():
+            gr.Markdown("### Buleyean-Trained (from the void)")
+            bule_out = gr.Textbox(label="buleyean-smollm2-360m", lines=12, interactive=False)
+
+    btn.click(compare, [prompt], [base_out, bule_out])
+    prompt.submit(compare, [prompt], [base_out, bule_out])
+
+    gr.Markdown("### Try these prompts:")
+    for p in ["hello", "How are you feeling today?", "I've been feeling really anxious lately.", "Write me a haiku about failure.", "What is the meaning of life?"]:
+        gr.Button(p, size="sm").click(
+            fn=lambda x=p: compare(x),
+            inputs=[],
+            outputs=[base_out, bule_out],
+        ).then(fn=lambda x=p: x, inputs=[], outputs=[prompt])
 
 if __name__ == "__main__":
     demo.launch(server_name="0.0.0.0", server_port=7860)

requirements.txt

@@ -1,7 +1,9 @@
-gradio>=5.0.0
+gradio>=4.44.0,<5.0.0
 transformers>=4.46.0
 peft>=0.13.0
 torch>=2.1.0
 accelerate>=1.0.0
 sentencepiece>=0.2.0
-huggingface-hub>=0.26.0
+protobuf>=4.25.0
+numpy>=1.26.0
+scipy>=1.12.0