Kenny Wang commited on
Commit
f463310
·
1 Parent(s): 219170b

Add Qwen3-8B v7 as default model (8.9 Mistral-strict)

Browse files
Files changed (1) hide show
  1. app.py +8 -4
app.py CHANGED
@@ -7,7 +7,11 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
7
  from peft import PeftModel
8
 
9
  MODELS = {
10
- "Qwen3-4B v7 (best-of-N, latest)": {
 
 
 
 
11
  "base": "Qwen/Qwen3-4B",
12
  "adapter": "hikewa/dialectic-qwen3-4b-v7-lora",
13
  },
@@ -122,15 +126,15 @@ demo = gr.ChatInterface(
122
  additional_inputs=[
123
  gr.Dropdown(
124
  choices=list(MODELS.keys()),
125
- value="Qwen3-4B v7 (best-of-N, latest)",
126
  label="Model",
127
  ),
128
  ],
129
  title="Dialectic Reasoning Models",
130
  description=(
131
  "Fine-tuned on dialectic reasoning traces with falsifiability-based quality filtering. "
132
- "v7 (113 best-of-N traces from 6 providers) scores 8.3 Mistral-strict with 13/14 strong verdicts. "
133
- "Pick a model and ask a question involving competing perspectives."
134
  ),
135
  examples=[
136
  ["Should AI systems be transparent about their reasoning, even when transparency reduces performance?"],
 
7
  from peft import PeftModel
8
 
9
  MODELS = {
10
+ "Qwen3-8B v7 (best-of-N, 8.9 strict)": {
11
+ "base": "Qwen/Qwen3-8B",
12
+ "adapter": "hikewa/dialectic-qwen3-8b-v7-lora",
13
+ },
14
+ "Qwen3-4B v7 (best-of-N, 8.3 strict)": {
15
  "base": "Qwen/Qwen3-4B",
16
  "adapter": "hikewa/dialectic-qwen3-4b-v7-lora",
17
  },
 
126
  additional_inputs=[
127
  gr.Dropdown(
128
  choices=list(MODELS.keys()),
129
+ value="Qwen3-8B v7 (best-of-N, 8.9 strict)",
130
  label="Model",
131
  ),
132
  ],
133
  title="Dialectic Reasoning Models",
134
  description=(
135
  "Fine-tuned on dialectic reasoning traces with falsifiability-based quality filtering. "
136
+ "8B v7 scores 8.9 Mistral-strict (up from 4B's 8.3) with deeper resolution and less hedging. "
137
+ "113 best-of-N traces from 6 providers. Pick a model and ask a question involving competing perspectives."
138
  ),
139
  examples=[
140
  ["Should AI systems be transparent about their reasoning, even when transparency reduces performance?"],