Tactical monochrome UI redesign and expanded training mission description

README.md

@@ -9,10 +9,12 @@ pinned: false
 
 # Math Conjecture Trainer Space
 
-This Space launches multi-stage `DeepSeek-Math` training for the conjecture
-project and pushes adapters to:
+This Space is the tactical training console for the project: it pulls released
+training corpus splits, builds runtime config from the SOTA curriculum YAML,
+executes multi-stage `DeepSeek-Math` fine-tuning, optionally evaluates
+self-consistency, and can publish adapters/checkpoints/training summaries to:
 
-- `NorthernTribe-Research/math-conjecture-model`
+- `NorthernTribe-Research/math-conjecture-model` (when push is enabled)
 
 ## What this Space does
 
@@ -41,7 +43,7 @@ automatically disabled for that run.
   `train_sota.py --dry-run`.
 - `Push Adapter to Hub`: controls whether `hub.push_to_hub` is enabled in the
   runtime config.
-- `Force Dataset Re-download`: bypasses cached local parquet files.
+- `Force Dataset Redownload`: bypasses cached local parquet files.
 - `Stop Active Run`: requests cancellation and terminates active subprocesses.
 - `Run Summary (JSON)`: structured output with config, status, and metrics.
 

app.py

@@ -44,6 +44,168 @@ CANCEL_REQUESTED = False
 ACTIVE_PROCESS: Optional[subprocess.Popen] = None
 ACTIVE_RUN_LABEL = ""
 
+TACTICAL_CSS = """
+@import url("https://fonts.googleapis.com/css2?family=IBM+Plex+Mono:wght@400;500;600&family=Rajdhani:wght@500;600;700&display=swap");
+
+:root {
+  --ops-bg: #070707;
+  --ops-bg-2: #0f0f0f;
+  --ops-panel: #111111;
+  --ops-panel-2: #161616;
+  --ops-border: #2a2a2a;
+  --ops-border-strong: #3d3d3d;
+  --ops-text: #ececec;
+  --ops-muted: #a8a8a8;
+  --ops-bright: #ffffff;
+}
+
+.gradio-container {
+  color: var(--ops-text) !important;
+  background:
+    linear-gradient(rgba(255, 255, 255, 0.02) 1px, transparent 1px),
+    linear-gradient(90deg, rgba(255, 255, 255, 0.02) 1px, transparent 1px),
+    radial-gradient(circle at 50% -10%, #1d1d1d 0%, #0f0f0f 38%, #070707 100%) !important;
+  background-size: 26px 26px, 26px 26px, 100% 100% !important;
+  font-family: "IBM Plex Mono", "JetBrains Mono", "Fira Code", monospace !important;
+}
+
+.gradio-container .prose h1,
+.gradio-container .prose h2,
+.gradio-container .prose h3,
+.gradio-container .prose p,
+.gradio-container .prose li,
+.gradio-container .prose strong {
+  color: var(--ops-text) !important;
+}
+
+.gradio-container .prose h1,
+.gradio-container .prose h2 {
+  font-family: "Rajdhani", "IBM Plex Mono", monospace !important;
+  letter-spacing: 0.08em !important;
+  text-transform: uppercase !important;
+}
+
+.gradio-container .prose code {
+  color: var(--ops-bright) !important;
+  background: #1b1b1b !important;
+  border: 1px solid var(--ops-border) !important;
+}
+
+.gradio-container .block,
+.gradio-container .form {
+  background: linear-gradient(180deg, var(--ops-panel) 0%, var(--ops-panel-2) 100%) !important;
+  border: 1px solid var(--ops-border) !important;
+  box-shadow: inset 0 0 0 1px rgba(255, 255, 255, 0.03), 0 12px 28px rgba(0, 0, 0, 0.35) !important;
+}
+
+.gradio-container label span,
+.gradio-container .block-info,
+.gradio-container [data-testid="block-info"] {
+  color: var(--ops-muted) !important;
+  letter-spacing: 0.12em !important;
+  text-transform: uppercase !important;
+  font-size: 0.74rem !important;
+}
+
+.gradio-container input,
+.gradio-container textarea {
+  background: #0c0c0c !important;
+  color: var(--ops-text) !important;
+  border: 1px solid var(--ops-border-strong) !important;
+  box-shadow: none !important;
+  font-family: "IBM Plex Mono", "JetBrains Mono", monospace !important;
+}
+
+.gradio-container input::placeholder,
+.gradio-container textarea::placeholder {
+  color: #7f7f7f !important;
+}
+
+.gradio-container input:focus,
+.gradio-container textarea:focus {
+  border-color: #656565 !important;
+  outline: none !important;
+}
+
+.gradio-container button {
+  border: 1px solid #565656 !important;
+  background: linear-gradient(180deg, #212121 0%, #151515 100%) !important;
+  color: var(--ops-bright) !important;
+  letter-spacing: 0.08em !important;
+  text-transform: uppercase !important;
+  font-family: "Rajdhani", "IBM Plex Mono", monospace !important;
+}
+
+.gradio-container button.primary,
+.gradio-container button.stop,
+.gradio-container button.secondary {
+  background: linear-gradient(180deg, #2a2a2a 0%, #171717 100%) !important;
+  border-color: #686868 !important;
+  color: #f7f7f7 !important;
+}
+
+.gradio-container button:hover {
+  filter: brightness(1.08);
+}
+
+.ops-header {
+  border: 1px solid var(--ops-border);
+  background: linear-gradient(180deg, #101010 0%, #0c0c0c 100%);
+  padding: 12px 14px;
+  margin: 2px 0 8px 0;
+}
+
+.ops-header-title {
+  font-family: "Rajdhani", "IBM Plex Mono", monospace;
+  letter-spacing: 0.16em;
+  text-transform: uppercase;
+  color: #f4f4f4;
+  font-weight: 700;
+  font-size: 1rem;
+}
+
+.ops-header-tags {
+  margin-top: 8px;
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+}
+
+.ops-tag {
+  border: 1px solid #474747;
+  background: #181818;
+  color: #d5d5d5;
+  padding: 3px 7px;
+  font-size: 0.72rem;
+  letter-spacing: 0.12em;
+  text-transform: uppercase;
+}
+"""
+
+TACTICAL_HEADER_HTML = """
+<div class="ops-header">
+  <div class="ops-header-title">Maths Conjecture Solutions // Training Operations Console</div>
+  <div class="ops-header-tags">
+    <span class="ops-tag">Tactical Monochrome</span>
+    <span class="ops-tag">Autonomous Auth</span>
+    <span class="ops-tag">Staged Curriculum</span>
+    <span class="ops-tag">Live Telemetry</span>
+  </div>
+</div>
+"""
+
+PROJECT_DESCRIPTION = """
+# Math Conjecture Trainer
+This console runs the full training operations lane for the `maths-conjuncture-solutions` project:
+
+1. Pull released parquet splits from `NorthernTribe-Research/math-conjecture-training-corpus`.
+2. Build runtime training configuration from `configs/deepseek_math_sota.yaml`.
+3. Execute multi-stage DeepSeek-Math curriculum fine-tuning via `scripts/train_sota.py`.
+4. Optionally evaluate adapters with pass@k-style sampling via `scripts/eval_sota.py`.
+5. Auto-resolve Hugging Face credentials, push adapters/checkpoints/summary when allowed, and stream live logs.
+6. Support preflight validation, abort control, cache strategy, and structured run-summary telemetry in one UI.
+"""
+
 
 def now_ts() -> str:
     return dt.datetime.utcnow().strftime("%Y-%m-%d %H:%M:%S UTC")
@@ -628,49 +790,47 @@ def run_pipeline(
 
 
 with gr.Blocks(title="Math Conjecture Trainer Space") as demo:
-    gr.Markdown(
-        """
-        # Math Conjecture Trainer
-        Launch multi-stage DeepSeek-Math fine-tuning on Space GPU and push adapters to your model repo.
-
-        Use **Preflight Only** to validate configuration and data flow without running a full train job.
-        HF authentication is resolved automatically from environment/credentials; no token input is required.
-        """
-    )
+    gr.HTML(TACTICAL_HEADER_HTML)
+    gr.Markdown(PROJECT_DESCRIPTION)
     with gr.Row():
         dataset_repo_id = gr.Textbox(
-            label="Dataset Repo",
+            label="Dataset Source",
             value="NorthernTribe-Research/math-conjecture-training-corpus",
         )
     with gr.Row():
         model_repo_id = gr.Textbox(
-            label="Model Repo",
+            label="Model Destination",
             value="NorthernTribe-Research/math-conjecture-model",
         )
         base_model_id = gr.Textbox(
-            label="Base Model",
+            label="Base Model ID",
             value="deepseek-ai/deepseek-math-v2",
         )
     with gr.Row():
-        start_stage = gr.Slider(label="Start Stage", minimum=1, maximum=3, step=1, value=1)
-        max_stages = gr.Slider(label="How Many Stages", minimum=1, maximum=3, step=1, value=3)
+        start_stage = gr.Slider(label="Stage Start", minimum=1, maximum=3, step=1, value=1)
+        max_stages = gr.Slider(label="Stage Count", minimum=1, maximum=3, step=1, value=3)
         run_eval = gr.Checkbox(label="Run Evaluation After Training", value=True)
     with gr.Row():
-        eval_k = gr.Slider(label="Eval K", minimum=1, maximum=8, step=1, value=4)
-        eval_samples = gr.Slider(label="Eval Max Samples", minimum=50, maximum=1000, step=50, value=300)
+        eval_k = gr.Slider(label="Evaluation K", minimum=1, maximum=8, step=1, value=4)
+        eval_samples = gr.Slider(label="Evaluation Max Samples", minimum=50, maximum=1000, step=50, value=300)
     with gr.Row():
         push_to_hub = gr.Checkbox(label="Push Adapter to Hub", value=True)
-        force_redownload = gr.Checkbox(label="Force Dataset Re-download", value=False)
+        force_redownload = gr.Checkbox(label="Force Dataset Redownload", value=False)
         preflight_only = gr.Checkbox(label="Preflight Only (No Training)", value=False)
 
     with gr.Row():
-        run_button = gr.Button("Start Train + Push", variant="primary")
-        stop_button = gr.Button("Stop Active Run", variant="stop")
-        clear_button = gr.Button("Clear Outputs")
-
-    status = gr.Textbox(label="Status", value="Idle", interactive=False)
-    logs = make_copyable_textbox(label="Logs", lines=24, max_lines=30, interactive=False)
-    run_summary = make_copyable_textbox(label="Run Summary (JSON)", lines=16, max_lines=24, interactive=False)
+        run_button = gr.Button("Execute Training Run", variant="primary")
+        stop_button = gr.Button("Abort Active Run", variant="stop")
+        clear_button = gr.Button("Reset Console")
+
+    status = gr.Textbox(label="Run Status", value="Idle", interactive=False)
+    logs = make_copyable_textbox(label="Telemetry Log", lines=24, max_lines=30, interactive=False)
+    run_summary = make_copyable_textbox(
+        label="Mission Summary (JSON)",
+        lines=16,
+        max_lines=24,
+        interactive=False,
+    )
 
     run_button.click(
         fn=run_pipeline,
@@ -694,4 +854,4 @@ with gr.Blocks(title="Math Conjecture Trainer Space") as demo:
 
 
 if __name__ == "__main__":
-    demo.queue(default_concurrency_limit=1).launch()
+    demo.queue(default_concurrency_limit=1).launch(css=TACTICAL_CSS)