Model size vs param count graph

app.py

@@ -16,6 +16,7 @@ from transformers import AutoTokenizer, AutoModelForCausalLM
 
 import gradio as gr
 import pandas as pd
+import plotly.graph_objects as go
 
 # ---------------------------------------------------------------------------
 # Logging
@@ -62,6 +63,21 @@ BASE_MODEL_DISPLAY: Dict[str, str] = {
     "OpenCerebral/Boris-1.3-75M": "Boris-1.3-75M",
 }
 
+MODEL_PARAMS: Dict[str, float] = {
+    "CodeSoft/MetaDiffusion-150M-ChatBase": 169.5e6,
+    "BananaMind/BananaMind-2-Medium-Chat": 49.6e6,
+    "SupraLabs/Supra2-100M-Instruct": 100.0e6,
+    "HuggingFaceTB/SmolLM2-135M-Instruct": 135.0e6,
+    "OpenCerebral/Boris-1.3-125M-Instruct": 125.0e6,
+    "OpenCerebral/Boris-1.3-75M-Instruct": 75.0e6,
+    "fromziro/Zero-v0.1-150M": 151.6e6,
+    "AxiomicLabs/GPT-X2.5-135M": 135.0e6,
+    "BananaMind/BananaMind-2-Pro": 139.0e6,
+    "HuggingFaceTB/SmolLM2-135M": 135.0e6,
+    "OpenCerebral/Boris-1.3-125M": 125.0e6,
+    "OpenCerebral/Boris-1.3-75M": 75.0e6,
+}
+
 FALLBACK_IDS: Dict[str, str] = {}
 
 MODERATION_MODEL_ID = "ifmain/ModerationBERT-En-02"
@@ -564,6 +580,44 @@ def leaderboard_dataframe(state: Optional[Dict[str, dict]] = None, spec: ArenaSp
     df.insert(0, "Rank", range(1, len(df) + 1))
     return df
 
+
+def elo_vs_params_fig(spec: ArenaSpec = MAIN_ARENA) -> "go.Figure":
+    """Scatter of ELO rating vs parameter count (log x) for an arena's models."""
+    state = load_elo(spec)
+    labels, xs, ys, texts, sizes = [], [], [], [], []
+    for mid in spec.model_ids:
+        p = MODEL_PARAMS.get(mid)
+        if not p:
+            continue
+        info = state.get(mid, {"rating": INIT_RATING, "battles": 0})
+        rating = float(info["rating"])
+        battles = int(info.get("battles", 0))
+        label = spec.display.get(mid, mid)
+        labels.append(label)
+        xs.append(p)
+        ys.append(rating)
+        texts.append(f"{label}<br>Params: {p/1e6:.1f}M<br>ELO: {rating:.1f}<br>Battles: {battles}")
+        sizes.append(max(8, min(30, 8 + battles * 1.5)))
+    fig = go.Figure()
+    fig.add_trace(go.Scatter(
+        x=xs, y=ys, mode="markers+text",
+        text=labels, textposition="top center",
+        textfont=dict(size=10, color="#cccccc"),
+        marker=dict(size=sizes, color=ys, colorscale="Viridis",
+                    line=dict(width=1, color="#ffffff")),
+        hovertext=texts, hoverinfo="text",
+    ))
+    fig.update_layout(
+        paper_bgcolor="rgba(0,0,0,0)", plot_bgcolor="rgba(0,0,0,0)",
+        font=dict(color="#e0e0e0"),
+        xaxis=dict(title="Parameters (log scale)", type="log",
+                   gridcolor="rgba(255,255,255,0.08)"),
+        yaxis=dict(title="ELO rating", gridcolor="rgba(255,255,255,0.08)"),
+        margin=dict(l=50, r=20, t=20, b=50),
+        showlegend=False,
+    )
+    return fig
+
 # ---------------------------------------------------------------------------
 # Chat logging to data/chats.jsonl
 # ---------------------------------------------------------------------------
@@ -970,8 +1024,11 @@ def create_demo() -> gr.Blocks:
                     column_widths=["5%", "15%", "25%", "12%", "7%", "7%", "7%", "7%", "7%"],
                     elem_id=elem_id,
                 )
+                gr.Markdown("### 📊 ELO vs Parameters")
+                gr.Markdown("X axis is log-scaled; bubble size scales with battle count.")
+                elo_params_plot = gr.Plot(value=elo_vs_params_fig(spec), show_label=False)
                 refresh_btn = gr.Button("🔄 Refresh", variant="secondary")
-                return {"tab": tab, "leaderboard": leaderboard, "refresh_btn": refresh_btn}
+                return {"tab": tab, "leaderboard": leaderboard, "elo_params_plot": elo_params_plot, "refresh_btn": refresh_btn}
 
         with gr.Tabs():
             ui = {}
@@ -1200,7 +1257,7 @@ def create_demo() -> gr.Blocks:
             )
 
         def on_refresh(spec: ArenaSpec):
-            return leaderboard_dataframe(load_elo(spec), spec)
+            return leaderboard_dataframe(load_elo(spec), spec), elo_vs_params_fig(spec)
 
         for spec in (MAIN_ARENA, BASE_ARENA):
             u, b = ui[spec.key], lb[spec.key]
@@ -1226,8 +1283,8 @@ def create_demo() -> gr.Blocks:
                 )
             u["new_round_btn"].click(fn=on_new_round, inputs=[], outputs=round_outputs)
             u["clear_btn"].click(fn=on_clear, inputs=[], outputs=[u["prompt"]] + round_outputs)
-            b["refresh_btn"].click(fn=lambda s=spec: on_refresh(s), inputs=[], outputs=[b["leaderboard"]])
-            b["tab"].select(fn=lambda s=spec: on_refresh(s), inputs=[], outputs=[b["leaderboard"]])
+            b["refresh_btn"].click(fn=lambda s=spec: on_refresh(s), inputs=[], outputs=[b["leaderboard"], b["elo_params_plot"]])
+            b["tab"].select(fn=lambda s=spec: on_refresh(s), inputs=[], outputs=[b["leaderboard"], b["elo_params_plot"]])
 
         gr.Markdown(
             "<div style=\"text-align:center; font-size:12px; opacity:0.65;\">"
@@ -1238,7 +1295,7 @@ def create_demo() -> gr.Blocks:
         )
 
         try:
-            demo.load(fn=lambda: on_refresh(MAIN_ARENA), inputs=[], outputs=[lb["main"]["leaderboard"]], show_progress="hidden")
+            demo.load(fn=lambda: on_refresh(MAIN_ARENA), inputs=[], outputs=[lb["main"]["leaderboard"], lb["main"]["elo_params_plot"]], show_progress="hidden")
         except Exception:
             pass