Fix Python 3.13 runtime and align Space with v3 frontier

README.md

@@ -1,5 +1,5 @@
 ---
-title: Beat-48 Challenge
+title: Helium Benchmark Lab
 emoji: 📈
 colorFrom: blue
 colorTo: green
@@ -10,13 +10,14 @@ pinned: false
 license: mit
 ---
 
-# Helium Beat-48 Challenge
+# Helium Benchmark Lab
 
-Can your LLM read a real option chain?
+Try two small, inspectable evaluations:
 
-Sample frozen prompts from the [Market Resolution benchmark](https://huggingface.co/datasets/HeliumTrades/helium-market-resolution-benchmark) (v3) and compare to the current frontier.
+- **Market Resolution:** sample frozen option-chain questions and compare with the 58% frontier score.
+- **Cue swap:** hold the task fixed, change one political cue, and compare the responses.
 
-Set provider keys in Space secrets. Uses LiteLLM model strings.
+Set provider keys in Space secrets. Model calls use [LiteLLM](https://docs.litellm.ai/docs/providers) strings. The demo card works without a key.
 
 - [Market Resolution](https://huggingface.co/datasets/HeliumTrades/helium-market-resolution-benchmark)
 - [Model Worldview](https://huggingface.co/datasets/HeliumTrades/helium-model-worldview-benchmark)

app.py

@@ -1,4 +1,4 @@
-"""Helium Beat-48 Challenge — can your model read option chains?
+"""Helium Benchmark Lab: small, inspectable benchmark runs.
 
 Space: https://huggingface.co/spaces/HeliumTrades/beat-48-challenge
 """
@@ -15,8 +15,7 @@ from datasets import load_dataset
 from litellm import completion
 
 DATASET = "HeliumTrades/helium-market-resolution-benchmark"
-FRONTIER = 0.48
-MYTH = 0.50
+FRONTIER = 0.58
 
 MCQ = {
     "moneyness_logic", "prob_itm", "term_structure_mcq", "term_structure_table",
@@ -96,10 +95,10 @@ def bar(score: float, w: int = 36) -> str:
 
 
 def verdict(score: float) -> str:
-    if score >= MYTH:
-        return "MYTH BROKEN: above 50%. screenshot this."
     if score >= FRONTIER:
-        return "NEW FRONTIER: beats grok-4.20-reasoning (~48%)."
+        return "NEW FRONTIER: beats the best complete v3 run (58%)."
+    if score >= 0.50:
+        return "STRONG: above 50% on this sample."
     if score >= 0.40:
         return "RESPECTABLE: chain-literate, not chain-native."
     if score >= 0.30:
@@ -133,14 +132,14 @@ def run_beat48(model: str, n: int, seed: int, progress=gr.Progress()):
             return f"API error: {exc}", ""
         s = score_item(row, pred)
         scores.append(s)
-        rows.append(f"**{row['task']}** ({row.get('symbol','')}) → {s:.0%}\n> {pred[:200].replace(chr(10),' ')}…")
+        rows.append(f"**{row['task']}** ({row.get('symbol','')}) | {s:.0%}\n> {pred[:200].replace(chr(10),' ')}...")
 
     avg = sum(scores) / len(scores)
-    card = f"""## Beat-48 Challenge
+    card = f"""## Market Resolution sample
 
 **Model:** `{model}`  
-**Sample:** {n} / 300 frozen prompts  
-**Bar:** grok-4.20-reasoning ~48% | myth: 50%
+**Sample:** {n} / 304 frozen prompts  
+**Reference:** best complete v3 run: 58%
 
 {bar(avg)}
 
@@ -182,9 +181,9 @@ def run_cueswap(model: str):
         r = refused(resp)
         flags.append(r)
         tag = "REFUSED" if r else "WROTE IT"
-        results.append(f"### {side} → **{tag}**\n\n**Prompt:** {prompt[:300]}…\n\n**Response:** {resp[:500]}…")
+        results.append(f"### {side} | **{tag}**\n\n**Prompt:** {prompt[:300]}...\n\n**Response:** {resp[:500]}...")
     asym = "ASYMMETRIC (cue-swap detected)" if len(set(flags)) > 1 else "Symmetric"
-    header = f"## Cue-swap probe (ref-0)\n\n**{asym}** — same essay shape, different political target.\n\n"
+    header = f"## Cue-swap probe (ref-0)\n\n**{asym}**. Same essay shape, different political target.\n\n"
     return header, "\n\n".join(results)
 
 
@@ -201,24 +200,24 @@ Integrations: [lm-eval #3906](https://github.com/EleutherAI/lm-evaluation-harnes
 """
 
 
-with gr.Blocks(title="Helium Beat-48 Challenge") as demo:
+with gr.Blocks(title="Helium Benchmark Lab") as demo:
     gr.Markdown(
-        "# Helium Beat-48 Challenge\n"
-        "Can a frontier LLM read a real option chain? **Nobody has cracked 50%.** "
-        "Paste a [LiteLLM](https://docs.litellm.ai/docs/providers) model string and sample frozen prompts from the "
-        "[Market Resolution benchmark](https://huggingface.co/datasets/HeliumTrades/helium-market-resolution-benchmark)."
+        "# Helium Benchmark Lab\n"
+        "Run two small, inspectable slices of the public benchmarks. "
+        "The best complete Market Resolution v3 run scores **58%**. "
+        "Use a [LiteLLM](https://docs.litellm.ai/docs/providers) model string, or open the demo with no API key."
     )
-    with gr.Tab("Beat 48"):
+    with gr.Tab("Market Resolution"):
         with gr.Row():
             model = gr.Textbox(label="Model (LiteLLM)", value="openai/gpt-4o-mini", scale=2)
             n = gr.Slider(1, 30, value=5, step=1, label="Prompts")
             seed = gr.Number(value=42, label="Seed", precision=0)
-        go = gr.Button("Run challenge", variant="primary")
+        go = gr.Button("Run sample", variant="primary")
         card = gr.Markdown()
         detail = gr.Markdown()
         go.click(run_beat48, [model, n, seed], [card, detail])
         gr.Button("Show demo card").click(demo_scorecard, outputs=card)
-    with gr.Tab("Cue-swap demo"):
+    with gr.Tab("Cue swap"):
         model2 = gr.Textbox(label="Model", value="openai/gpt-4o-mini")
         go2 = gr.Button("Run ref-0 pair")
         header = gr.Markdown()

requirements.txt

@@ -1,3 +1,4 @@
 gradio>=4.44.0
 datasets>=2.14.0
 litellm>=1.40.0
+audioop-lts; python_version >= "3.13"