Upload app.py with huggingface_hub

app.py

@@ -7,53 +7,22 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import PeftModel
 
 MODELS = {
-    "Qwen3-8B v7 (best-of-N, 8.9 strict)": {
+    "Qwen3-8B v8 (best)": {
         "base": "Qwen/Qwen3-8B",
-        "adapter": "hikewa/dialectic-qwen3-8b-v7-lora",
+        "adapter": "hikewa/dialectic-qwen3-8b-v8-lora",
     },
-    "Qwen3-4B v7 (best-of-N, 8.3 strict)": {
+    "Qwen3-4B v8": {
         "base": "Qwen/Qwen3-4B",
-        "adapter": "hikewa/dialectic-qwen3-4b-v7-lora",
-    },
-    "Qwen3-4B v6 (falsifiable)": {
-        "base": "Qwen/Qwen3-4B",
-        "adapter": "hikewa/dialectic-qwen3-4b-v6-lora",
-    },
-    "Qwen3-4B v3": {
-        "base": "Qwen/Qwen3-4B",
-        "adapter": "hikewa/dialectic-qwen3-4b-v3-lora",
-    },
-    "Qwen3-4B v4 (axiom)": {
-        "base": "Qwen/Qwen3-4B",
-        "adapter": "hikewa/dialectic-qwen3-4b-v4-lora",
-    },
-    "Qwen3-8B v3": {
-        "base": "Qwen/Qwen3-8B",
-        "adapter": "hikewa/dialectic-qwen3-8b-v3-lora",
-    },
-    "Qwen3-8B v1": {
-        "base": "Qwen/Qwen3-8B",
-        "adapter": "hikewa/dialectic-qwen3-8b-lora",
-    },
-    "Qwen3-4B v2": {
-        "base": "Qwen/Qwen3-4B",
-        "adapter": "hikewa/dialectic-qwen3-4b-v2-lora",
-    },
-    "Qwen2.5-1.5B": {
-        "base": "Qwen/Qwen2.5-1.5B-Instruct",
-        "adapter": "hikewa/dialectic-qwen2.5-1.5b-lora",
-    },
-    "Qwen2.5-0.5B": {
-        "base": "Qwen/Qwen2.5-0.5B-Instruct",
-        "adapter": "hikewa/dialectic-qwen2.5-0.5b-lora",
+        "adapter": "hikewa/dialectic-qwen3-4b-v8-lora",
     },
 }
 
 SYSTEM_PROMPT = (
-    "You reason carefully through problems by considering competing "
-    "perspectives before reaching a conclusion. You identify genuine "
-    "tensions, engage with the strongest form of each argument, and "
-    "integrate insights rather than picking sides or hedging."
+    "You reason carefully through problems by considering competing perspectives "
+    "before reaching a conclusion. You identify genuine tensions, engage with "
+    "the strongest form of each argument, and integrate insights rather than "
+    "picking sides or hedging. Be specific through mechanisms and conditions, "
+    "not through citations or statistics."
 )
 
 loaded = {"name": None, "model": None, "tokenizer": None}
@@ -109,7 +78,7 @@ def respond(message, history, model_name):
     with torch.no_grad():
         outputs = model.generate(
             **inputs,
-            max_new_tokens=1024,
+            max_new_tokens=512,
             temperature=0.7,
             do_sample=True,
             repetition_penalty=1.1,
@@ -126,15 +95,15 @@ demo = gr.ChatInterface(
     additional_inputs=[
         gr.Dropdown(
             choices=list(MODELS.keys()),
-            value="Qwen3-8B v7 (best-of-N, 8.9 strict)",
+            value="Qwen3-8B v8 (best)",
             label="Model",
         ),
     ],
     title="Dialectic Reasoning Models",
     description=(
-        "Fine-tuned on dialectic reasoning traces with falsifiability-based quality filtering. "
-        "8B v7 scores 8.9 Mistral-strict (up from 4B's 8.3) with deeper resolution and less hedging. "
-        "113 best-of-N traces from 6 providers. Pick a model and ask a question involving competing perspectives."
+        "Fine-tuned on 212 grounded dialectic reasoning traces (v8). "
+        "Mechanism-based specificity, no fabricated citations. "
+        "Pick a model size and ask a question involving competing perspectives."
     ),
     examples=[
         ["Should AI systems be transparent about their reasoning, even when transparency reduces performance?"],