feat: add personality models tab and all models catalog

Three tabs: Compare (live inference), Personality Models (sweep results
and 32B downloads), All Models (full catalog with links).
Five personality-modulated 32B adapters now linked.

app.py

@@ -4,61 +4,63 @@ LIVE inference only. Every response generated in real-time.
 """
 
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
+from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import PeftModel
 import torch
-import os
 
-print("Loading base model (SmolLM2-360M-Instruct)...", flush=True)
-base_model_id = "HuggingFaceTB/SmolLM2-360M-Instruct"
-buleyean_adapter = "forkjoin-ai/buleyean-smollm2-360m"
+BASE_MODEL_ID = "HuggingFaceTB/SmolLM2-360M-Instruct"
+BULEYEAN_ADAPTER = "forkjoin-ai/buleyean-smollm2-360m"
 
-tokenizer = AutoTokenizer.from_pretrained(base_model_id)
+print("Loading tokenizer...", flush=True)
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
 if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token
 
-# Load base model
+print(f"Loading base model ({BASE_MODEL_ID})...", flush=True)
 base_model = AutoModelForCausalLM.from_pretrained(
-    base_model_id,
+    BASE_MODEL_ID,
     torch_dtype=torch.float32,
     device_map="cpu",
     trust_remote_code=True,
 )
 print("Base model loaded.", flush=True)
 
-# Load Buleyean model (base + LoRA adapter)
 print("Loading Buleyean adapter...", flush=True)
 bule_base = AutoModelForCausalLM.from_pretrained(
-    base_model_id,
+    BASE_MODEL_ID,
     torch_dtype=torch.float32,
     device_map="cpu",
     trust_remote_code=True,
 )
 try:
-    bule_model = PeftModel.from_pretrained(bule_base, buleyean_adapter)
+    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("Both models ready. Live inference active.", flush=True)
+print("All models ready. Live inference active.", flush=True)
 
 
-def generate(prompt, model):
+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)
+    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=300,
-            temperature=0.7,
-            top_p=0.9,
+            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)
+    response = tokenizer.decode(
+        outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True
+    )
     return response.strip()
 
 
@@ -70,38 +72,151 @@ def compare(prompt):
     return base_out, bule_out
 
 
-with gr.Blocks(title="The Void", theme=gr.themes.Base(primary_hue="indigo")) as demo:
+# ---------------------------------------------------------------------------
+# UI
+# ---------------------------------------------------------------------------
+
+with gr.Blocks(
+    title="The Void -- Buleyean RL",
+    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.**
 
-Type any prompt. Both models run inference right now on this machine.
+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
+""")
+
+    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) |
+**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()
+```
+""")
+
+    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) |
 
-Base: [SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct)
-Buleyean: [buleyean-smollm2-360m](https://huggingface.co/forkjoin-ai/buleyean-smollm2-360m) -- same model, trained from rejection alone
+### Personality-Modulated (32B)
 
-[Library](https://github.com/forkjoin-ai/buleyean-rl) | [Paper](https://huggingface.co/forkjoin-ai) | 500+ Lean 4 theorems, zero sorry
+| 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)
 """)
-    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])
+
+    gr.Markdown("""---
+*Built with [Buleyean RL](https://github.com/forkjoin-ai/buleyean-rl). The void is where the information is.*
+""")
+
 
 if __name__ == "__main__":
     demo.launch(server_name="0.0.0.0", server_port=7860)