Upload folder using huggingface_hub

demo/ui.py

@@ -850,6 +850,90 @@ footer, .gradio-container footer { display: none !important; }
     margin-top: -0.5rem !important;
 }
 
+/* Episode-done dashboard: KPI row with big metric cards */
+.gradio-container .lf-done {
+    background: linear-gradient(180deg,
+        rgba(226,135,99,0.06) 0%,
+        rgba(42,40,36,0) 60%);
+    border: 1px solid var(--lf-border);
+    border-radius: 12px;
+    padding: 1.2rem 1.25rem;
+    margin: 1.1rem 0 0.6rem;
+}
+.gradio-container .lf-done-head {
+    display: flex; align-items: baseline; gap: 0.85rem;
+    margin-bottom: 0.9rem;
+}
+.gradio-container .lf-done-flag {
+    color: var(--lf-accent);
+    font-family: "Inter", sans-serif;
+    font-weight: 600; font-size: 0.75rem;
+    letter-spacing: 0.11em; text-transform: uppercase;
+    padding: 0.15rem 0.55rem;
+    border: 1px solid var(--lf-accent);
+    border-radius: 5px;
+}
+.gradio-container .lf-done-reason {
+    color: var(--lf-text-subtle);
+    font-size: 0.84rem;
+}
+.gradio-container .lf-done-reason code {
+    font-family: "JetBrains Mono", monospace;
+    background: transparent !important;
+    border: none !important;
+    color: var(--lf-text-muted) !important;
+    padding: 0 !important;
+}
+
+.gradio-container .lf-kpi-row {
+    display: grid;
+    grid-template-columns: repeat(3, 1fr);
+    gap: 0.8rem;
+}
+.gradio-container .lf-kpi {
+    background: var(--lf-surface-alt);
+    border: 1px solid var(--lf-border-soft);
+    border-radius: 10px;
+    padding: 0.9rem 1rem;
+    min-width: 0;
+}
+.gradio-container .lf-kpi-label {
+    color: var(--lf-text-subtle);
+    font-family: "Inter", sans-serif;
+    font-size: 0.7rem;
+    font-weight: 600;
+    letter-spacing: 0.1em;
+    text-transform: uppercase;
+    margin-bottom: 0.35rem;
+}
+.gradio-container .lf-kpi-value {
+    font-family: "Source Serif 4", Georgia, serif;
+    font-weight: 500;
+    font-size: 1.9rem;
+    color: var(--lf-text);
+    letter-spacing: -0.025em;
+    line-height: 1.1;
+}
+.gradio-container .lf-kpi-sub {
+    color: var(--lf-text-subtle);
+    font-size: 0.72rem;
+    margin-top: 0.3rem;
+    font-family: "JetBrains Mono", monospace;
+}
+.gradio-container .lf-kpi-good .lf-kpi-value { color: #7ab68c; }
+.gradio-container .lf-kpi-warn .lf-kpi-value { color: #e4b264; }
+.gradio-container .lf-kpi-bad  .lf-kpi-value { color: #d47d6a; }
+.gradio-container .lf-kpi-good { border-color: rgba(122,182,140,0.35); }
+.gradio-container .lf-kpi-warn { border-color: rgba(228,178,100,0.35); }
+.gradio-container .lf-kpi-bad  { border-color: rgba(212,125,106,0.35); }
+
+/* Responsive: stack KPIs on narrow */
+@media (max-width: 720px) {
+    .gradio-container .lf-kpi-row {
+        grid-template-columns: 1fr;
+    }
+}
+
 /* Main pane plots+outputs */
 .gradio-container .gr-plot, .gradio-container .plot-wrap {
     background: var(--lf-surface-alt) !important;
@@ -1539,17 +1623,50 @@ def _llm_auto_run(endpoint_choice, custom_url, api_key, model_name,
 
         if obs.done:
             bk = obs.r_optcoder_breakdown or {}
-            log_lines.extend([
-                f"---",
-                f"### 🏁 Episode done",
-                f"- Reason: `{(obs.last_action_result or {}).get('reason')}`",
-                f"- **Terminal reward: `{obs.r_optcoder:+.3f}`**",
-                f"- Final Adam-shortfall: `{obs.final_regret:.3f}`",
-                "",
-                f"| component | value |",
-                f"|---|---|",
-                *[f"| `{k}` | `{v:+.4f}` " for k, v in bk.items()],
-            ])
+            reward_val = obs.r_optcoder or 0.0
+            my_prog = bk.get("my_progress", 0.0)
+            adam_prog = bk.get("adam_progress", 0.0)
+            speedup = bk.get("speedup_vs_adam", 0.0)
+            reason = (obs.last_action_result or {}).get("reason", "?")
+
+            # Tone of the reward KPI — green if positive, red if negative
+            reward_tone = ("lf-kpi-good" if reward_val >= 0.5 else
+                           ("lf-kpi-warn" if reward_val >= 0 else "lf-kpi-bad"))
+            speedup_display = (f"{speedup:.2f}×" if speedup < 100
+                               else f"{speedup:.0f}×")
+            speedup_tone = ("lf-kpi-good" if speedup >= 1.0
+                            else "lf-kpi-warn")
+
+            episode_done_html = (
+                "<div class='lf-done'>"
+                "  <div class='lf-done-head'>"
+                "    <span class='lf-done-flag'>Episode complete</span>"
+                f"   <span class='lf-done-reason'>ended by "
+                f"<code>{reason}</code></span>"
+                "  </div>"
+                "  <div class='lf-kpi-row'>"
+                "    <div class='lf-kpi " f"{reward_tone}" "'>"
+                "      <div class='lf-kpi-label'>Terminal reward</div>"
+                f"     <div class='lf-kpi-value'>{reward_val:+.3f}</div>"
+                "      <div class='lf-kpi-sub'>GRPO training scalar</div>"
+                "    </div>"
+                "    <div class='lf-kpi " f"{speedup_tone}" "'>"
+                "      <div class='lf-kpi-label'>Speedup vs tuned Adam</div>"
+                f"     <div class='lf-kpi-value'>{speedup_display}</div>"
+                f"     <div class='lf-kpi-sub'>my {my_prog:.3g} · "
+                f"adam {adam_prog:.3g}</div>"
+                "    </div>"
+                "    <div class='lf-kpi'>"
+                "      <div class='lf-kpi-label'>Adam shortfall</div>"
+                f"     <div class='lf-kpi-value'>{obs.final_regret:.3f}</div>"
+                "      <div class='lf-kpi-sub'>0 = matched/beat Adam</div>"
+                "    </div>"
+                "  </div>"
+                "</div>"
+            )
+
+            log_lines.extend([f"", episode_done_html, f""])
+
             reward_plot = _reward_breakdown_plot({
                 "r_regret":      bk.get("r_regret", 0),
                 "r_convergence": bk.get("r_convergence", 0),
@@ -1557,7 +1674,7 @@ def _llm_auto_run(endpoint_choice, custom_url, api_key, model_name,
                 "r_novelty":     bk.get("r_novelty", 0),
                 "-r_budget":     -bk.get("r_budget", 0),
                 "-r_eval_fail":  -bk.get("r_eval_failures", 0),
-            }, obs.r_optcoder or 0.0)
+            }, reward_val)
             yield ("\n".join(log_lines),
                    obs.model_dump(exclude_none=True),
                    reward_plot)