Update app.py

app.py

@@ -393,7 +393,7 @@ def _diff_generate_response(model, tokenizer, prompt_ids, gen_len, num_steps, te
 # ELO persistence
 # ---------------------------------------------------------------------------
 def init_elo_state() -> Dict[str, dict]:
-    return {mid: {"rating": float(INIT_RATING), "wins": 0, "losses": 0, "battles": 0, "ties": 0} for mid in MODEL_IDS}
+    return {mid: {"rating": float(INIT_RATING), "wins": 0, "losses": 0, "battles": 0, "ties": 0, "both_bad": 0} for mid in MODEL_IDS}
 
 def load_elo() -> Dict[str, dict]:
     if get_elo_file().exists():
@@ -402,7 +402,14 @@ def load_elo() -> Dict[str, dict]:
                 data = json.load(f)
             for mid in MODEL_IDS:
                 if mid not in data:
-                    data[mid] = {"rating": float(INIT_RATING), "wins": 0, "losses": 0, "battles": 0, "ties": 0}
+                    data[mid] = {"rating": float(INIT_RATING), "wins": 0, "losses": 0, "battles": 0, "ties": 0, "both_bad": 0}
+                else:
+                    data[mid].setdefault("rating", float(INIT_RATING))
+                    data[mid].setdefault("wins", 0)
+                    data[mid].setdefault("losses", 0)
+                    data[mid].setdefault("battles", 0)
+                    data[mid].setdefault("ties", 0)
+                    data[mid].setdefault("both_bad", 0)
             return data
         except Exception as e:
             logger.warning(f"Failed to load ELO file: {e}, resetting")
@@ -431,7 +438,7 @@ def update_elo(state: Dict[str, dict], model_a: str, model_b: str, winner: Optio
         sa = 1.0
     elif winner == model_b:
         sa = 0.0
-    elif winner is None or winner == "tie":
+    elif winner is None or winner == "tie" or winner == "both_bad":
         sa = 0.5
     else:
         raise ValueError(f"Unexpected winner: {winner}")
@@ -446,9 +453,14 @@ def update_elo(state: Dict[str, dict], model_a: str, model_b: str, winner: Optio
     elif sa == 0.0:
         state[model_b]["wins"] += 1
         state[model_a]["losses"] += 1
-    else:
+    elif winner != "both_bad":
         state[model_a]["ties"] += 1
         state[model_b]["ties"] += 1
+
+    if winner == "both_bad":
+        state[model_a]["both_bad"] = state[model_a].get("both_bad", 0) + 1
+        state[model_b]["both_bad"] = state[model_b].get("both_bad", 0) + 1
+
     save_elo(state)
     return state
 
@@ -466,6 +478,7 @@ def leaderboard_dataframe(state: Optional[Dict[str, dict]] = None) -> pd.DataFra
             "Wins": int(info["wins"]),
             "Losses": int(info["losses"]),
             "Ties": int(info.get("ties", 0)),
+            "Both Bad": int(info.get("both_bad", 0)),
         })
     df = pd.DataFrame(rows)
     df = df.sort_values(by="ELO", ascending=False).reset_index(drop=True)
@@ -478,7 +491,7 @@ def leaderboard_dataframe(state: Optional[Dict[str, dict]] = None) -> pd.DataFra
 def log_battle(prompt: str, model_a: str, model_b: str, response_a: str, response_b: str, chosen: str, winner_model: str):
     """
     Append one battle record to data/chats.jsonl.
-    Fields: prompt, response_a, response_b, model_a, model_b, chosen (A/B), winner_model, timestamp
+    Fields: prompt, response_a, response_b, model_a, model_b, chosen (A/B/tie/both_bad), winner_model, timestamp
     Spec: keeps user's message, two responses, each model's names, and what response user chose.
     """
     try:
@@ -490,7 +503,7 @@ def log_battle(prompt: str, model_a: str, model_b: str, response_a: str, respons
             "model_b": model_b,
             "response_a": response_a,
             "response_b": response_b,
-            "chosen": chosen,  # "A" / "B" / "tie"
+            "chosen": chosen,  # "A" / "B" / "tie" / "both_bad"
             "winner_model": winner_model,
             "chosen_response": response_a if chosen == "A" else response_b if chosen == "B" else "",
         }
@@ -831,7 +844,8 @@ def create_demo() -> gr.Blocks:
                 with gr.Row():
                     vote_a = gr.Button("👈 Vote for A", variant="secondary", interactive=False, elem_classes=["vote-btn"])
                     vote_tie = gr.Button("🤝 Tie", variant="secondary", interactive=False, elem_classes=["vote-btn"])
-                    vote_b = gr.Button("Vote for B", variant="secondary", interactive=False, elem_classes=["vote-btn"])
+                    vote_both_bad = gr.Button("👎 Both Bad", variant="secondary", interactive=False, elem_classes=["vote-btn"])
+                    vote_b = gr.Button("Vote for B 👉", variant="secondary", interactive=False, elem_classes=["vote-btn"])
 
                 status = gr.Markdown(visible=False)
                 new_round_btn = gr.Button("🔄 New Round", visible=False, variant="secondary")
@@ -846,11 +860,11 @@ def create_demo() -> gr.Blocks:
                 gr.Markdown("### 🏆 ELO Leaderboard")
                 leaderboard = gr.Dataframe(
                     value=df_init,
-                    headers=["Rank", "Model", "Model ID", "ELO", "Battles", "Wins", "Losses", "Ties"],
-                    datatype=["number", "str", "str", "number", "number", "number", "number", "number"],
+                    headers=["Rank", "Model", "Model ID", "ELO", "Battles", "Wins", "Losses", "Ties", "Both Bad"],
+                    datatype=["number", "str", "str", "number", "number", "number", "number", "number", "number"],
                     interactive=False,
                     wrap=False,
-                    column_widths=["5%", "20%", "35%", "12%", "7%", "7%", "7%", "7%"],
+                    column_widths=["5%", "15%", "25%", "12%", "7%", "7%", "7%", "7%", "7%"],
                     elem_id="leaderboard",
                 )
                 with gr.Row():
@@ -865,13 +879,13 @@ def create_demo() -> gr.Blocks:
                 return (
                     gr.update(value="", placeholder="Please enter a prompt first!"),
                     gr.update(value=""),
-                    gr.update(value=""),
                     gr.update(visible=False),
                     gr.update(visible=False),
                     gr.update(visible=False, value=""),
                     gr.update(interactive=False),
                     gr.update(interactive=False),
                     gr.update(interactive=False),
+                    gr.update(interactive=False),
                     gr.update(visible=False),
                     "", "", False, user_prompt, last_pair_val,
                     leaderboard_dataframe(load_elo())
@@ -895,6 +909,7 @@ def create_demo() -> gr.Blocks:
                 gr.update(interactive=True),
                 gr.update(interactive=True),
                 gr.update(interactive=True),
+                gr.update(interactive=True),
                 gr.update(visible=False),
                 a, b, False, user_prompt, (a, b),
                 leaderboard_dataframe(load_elo())
@@ -909,6 +924,7 @@ def create_demo() -> gr.Blocks:
                     gr.update(interactive=False),
                     gr.update(interactive=False),
                     gr.update(interactive=False),
+                    gr.update(interactive=False),
                     gr.update(visible=False),
                     voted,
                     leaderboard_dataframe(load_elo())
@@ -925,6 +941,10 @@ def create_demo() -> gr.Blocks:
                 winner = None
                 win_label = "Tie"
                 chosen = "tie"
+            elif choice == "Both Bad":
+                winner = "both_bad"
+                win_label = "Both Bad"
+                chosen = "both_bad"
             else:
                 winner = model_b
                 win_label = "B"
@@ -945,6 +965,12 @@ def create_demo() -> gr.Blocks:
                     f"**ELO update:** {MODEL_DISPLAY.get(model_a, model_a)} {ra_before:.1f} → {ra_after:.1f} ({delta_a:+.1f}) | "
                     f"{MODEL_DISPLAY.get(model_b, model_b)} {rb_before:.1f} → {rb_after:.1f} ({delta_b:+.1f})"
                 )
+            elif choice == "Both Bad":
+                status_text = (
+                    f"You voted **Both Bad**: no winner\n\n"
+                    f"**ELO update:** {MODEL_DISPLAY.get(model_a, model_a)} {ra_before:.1f} → {ra_after:.1f} ({delta_a:+.1f}) | "
+                    f"{MODEL_DISPLAY.get(model_b, model_b)} {rb_before:.1f} → {rb_after:.1f} ({delta_b:+.1f})"
+                )
             else:
                 status_text = (
                     f"You voted **{win_label}**: the winner is `{winner}`\n\n"
@@ -961,6 +987,7 @@ def create_demo() -> gr.Blocks:
                 gr.update(interactive=False),
                 gr.update(interactive=False),
                 gr.update(interactive=False),
+                gr.update(interactive=False),
                 gr.update(visible=True),
                 True,
                 df
@@ -976,6 +1003,7 @@ def create_demo() -> gr.Blocks:
                 gr.update(interactive=False),
                 gr.update(interactive=False),
                 gr.update(interactive=False),
+                gr.update(interactive=False),
                 gr.update(visible=False),
                 "", "", False, ""
             )
@@ -991,6 +1019,7 @@ def create_demo() -> gr.Blocks:
                 gr.update(interactive=False),
                 gr.update(interactive=False),
                 gr.update(interactive=False),
+                gr.update(interactive=False),
                 gr.update(visible=False),
                 "", "", False, ""
             )
@@ -1001,40 +1030,45 @@ def create_demo() -> gr.Blocks:
         submit_btn.click(
             fn=on_submit,
             inputs=[prompt, last_pair],
-            outputs=[response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state, last_pair, leaderboard],
+            outputs=[response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state, last_pair, leaderboard],
         )
 
         prompt.submit(
             fn=on_submit,
             inputs=[prompt, last_pair],
-            outputs=[response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state, last_pair, leaderboard],
+            outputs=[response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state, last_pair, leaderboard],
         )
 
         vote_a.click(
             fn=lambda ma, mb, ra, rb, pr, vd: on_vote("A", ma, mb, ra, rb, pr, vd),
             inputs=[model_a_state, model_b_state, response_a, response_b, prompt_state, voted_state],
-            outputs=[reveal_a, reveal_b, status, vote_a, vote_tie, vote_b, new_round_btn, voted_state, leaderboard],
+            outputs=[reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, voted_state, leaderboard],
         )
         vote_tie.click(
             fn=lambda ma, mb, ra, rb, pr, vd: on_vote("Tie", ma, mb, ra, rb, pr, vd),
             inputs=[model_a_state, model_b_state, response_a, response_b, prompt_state, voted_state],
-            outputs=[reveal_a, reveal_b, status, vote_a, vote_tie, vote_b, new_round_btn, voted_state, leaderboard],
+            outputs=[reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, voted_state, leaderboard],
+        )
+        vote_both_bad.click(
+            fn=lambda ma, mb, ra, rb, pr, vd: on_vote("Both Bad", ma, mb, ra, rb, pr, vd),
+            inputs=[model_a_state, model_b_state, response_a, response_b, prompt_state, voted_state],
+            outputs=[reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, voted_state, leaderboard],
         )
         vote_b.click(
             fn=lambda ma, mb, ra, rb, pr, vd: on_vote("B", ma, mb, ra, rb, pr, vd),
             inputs=[model_a_state, model_b_state, response_a, response_b, prompt_state, voted_state],
-            outputs=[reveal_a, reveal_b, status, vote_a, vote_tie, vote_b, new_round_btn, voted_state, leaderboard],
+            outputs=[reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, voted_state, leaderboard],
         )
 
         new_round_btn.click(
             fn=on_new_round,
             inputs=[],
-            outputs=[response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state],
+            outputs=[response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state],
         )
         clear_btn.click(
             fn=on_clear,
             inputs=[],
-            outputs=[prompt, response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state],
+            outputs=[prompt, response_a, response_b, reveal_a, reveal_b, status, vote_a, vote_tie, vote_both_bad, vote_b, new_round_btn, model_a_state, model_b_state, voted_state, prompt_state],
         )
 
         refresh_btn.click(fn=on_refresh, inputs=[], outputs=[leaderboard])