Upload folder using huggingface_hub

server/env/__init__.py

@@ -1,4 +1,10 @@
-from .environment import ContentGuardEnv
 from .tasks import TASKS
 
 __all__ = ["ContentGuardEnv", "TASKS"]
+
+
+def __getattr__(name):
+	if name == "ContentGuardEnv":
+		from .environment import ContentGuardEnv
+		return ContentGuardEnv
+	raise AttributeError(f"module '{__name__}' has no attribute '{name}'")

server/env/data_gen.py

@@ -4,26 +4,52 @@ Synthetic & Hugging Face content case generator for ContentGuardEnv.
 Generates realistic social media posts with associated metadata,
 user account history, and ground-truth moderation decisions.
 Integrates natively with Hugging Face datasets.
+
+By default, external dataset loading is disabled to keep startup/test
+paths deterministic. Enable with CONTENT_GUARD_USE_HF_DATA=1.
 """
 
+import os
 import random
 import copy
-from typing import Any, Dict, List, Tuple
+from typing import Any, Dict, Tuple
 import logging
 
 logger = logging.getLogger(__name__)
 
-# Attempt to load Real World Data from Hugging Face
-HF_DATA = None
+HF_DATASET = None
+_HF_DATA_ATTEMPTED = False
+_HF_DATA_ENABLED = os.getenv("CONTENT_GUARD_USE_HF_DATA", "0").strip().lower() in {"1", "true", "yes", "on"}
+
 try:
-    from datasets import load_dataset
-    logger.info("Initializing Hugging Face Datasets connection...")
-    # Load a small, highly relevant dataset of flagged conversational text
-    hf_dataset = load_dataset("SetFit/toxic_conversations", split="train")
-    HF_DATA = list(hf_dataset)
-    logger.info(f"Loaded {len(HF_DATA)} real evaluation cases from Hugging Face!")
-except Exception as e:
-    logger.warning(f"Could not load Hugging Face dataset (will use synthetic): {str(e)}")
+    from datasets import load_dataset as _hf_load_dataset
+except Exception:
+    _hf_load_dataset = None
+
+
+def _maybe_load_hf_data() -> None:
+    """Lazy one-time dataset initialization to avoid import-time blocking."""
+    global HF_DATASET, _HF_DATA_ATTEMPTED
+
+    if _HF_DATA_ATTEMPTED:
+        return
+    _HF_DATA_ATTEMPTED = True
+
+    if not _HF_DATA_ENABLED:
+        logger.info("Hugging Face dataset disabled (CONTENT_GUARD_USE_HF_DATA=0). Using synthetic cases.")
+        return
+    if _hf_load_dataset is None:
+        logger.warning("datasets package unavailable; using synthetic cases.")
+        return
+
+    try:
+        logger.info("Initializing Hugging Face Datasets connection...")
+        # Keep dataset bounded for memory and faster cold starts.
+        HF_DATASET = _hf_load_dataset("SetFit/toxic_conversations", split="train[:1000]")
+        logger.info(f"Loaded {len(HF_DATASET)} Hugging Face evaluation records.")
+    except Exception as e:
+        logger.warning(f"Could not load Hugging Face dataset (falling back to synthetic): {str(e)}")
+        HF_DATASET = None
 
 
 # ── Post templates (Fallback / Structure) ──────────────────────────────────────
@@ -83,7 +109,10 @@ DEVICE_TYPES = ["mobile_ios", "mobile_android", "desktop_web", "desktop_app"]
 
 def get_real_hf_case() -> Tuple[str, str, int, str]:
     """Pulls a random record from the real Hugging Face dataset."""
-    record = random.choice(HF_DATA)
+    if HF_DATASET is None or len(HF_DATASET) == 0:
+        raise ValueError("HF dataset is not initialized.")
+
+    record = HF_DATASET[random.randrange(len(HF_DATASET))]
     # SetFit/toxic_conversations uses 'text' and 'label' (0=safe, 1=toxic)
     content = record.get("text", "")
     is_toxic = record.get("label", 0) == 1
@@ -105,10 +134,11 @@ def generate_case(task_id: str) -> Tuple[Dict[str, Any], Dict[str, Any]]:
     """Generate a content moderation case leveraging real HF data if available."""
     
     # 1. Base Meta structure
-    template = random.choice(POSTS)
+    template = copy.deepcopy(random.choice(POSTS))
+    _maybe_load_hf_data()
     
     # 2. Content & Ground Truth assignment
-    if HF_DATA:
+    if HF_DATASET is not None and len(HF_DATASET) > 0:
         # Real AI Training Data Path
         content, violation, severity, action = get_real_hf_case()
         template["violation"] = violation

server/static/app.js

@@ -84,9 +84,7 @@ const app = {
             return;
         }
 
-        this.closeReward(true);
-        this.terminalPrint('LOG: Quick-cycle trigger received (/). Starting a new episode...');
-        this.startEpisode(this.currentTask);
+        this.closeReward(true, true);
     },
 
     connectWS: function() {
@@ -375,7 +373,7 @@ const app = {
         if (!this.scrollPending) {
             this.scrollPending = true;
             requestAnimationFrame(() => {
-                this.scrollOversightToBottom(false);
+                this.scrollOversightToBottom(true);
                 this.scrollPending = false;
             });
         }
@@ -445,8 +443,16 @@ const app = {
         }
     },
     
-    closeReward: function(silent) {
+    closeReward: function(silent, autoStartNext) {
         document.getElementById('reward-overlay').style.display = 'none';
+
+        const shouldAutoStart = !!this.currentTask && (autoStartNext === true || (!this.isAutoTraining && this.episodeDone));
+        if (shouldAutoStart) {
+            this.terminalPrint('LOG: Dismiss received. Starting next episode...');
+            this.startEpisode(this.currentTask);
+            return;
+        }
+
         if (!silent && !this.isAutoTraining) {
             this.terminalPrint(`LOG: Alignment evaluation captured and dismissed.`);
         }

server/static/style.css

@@ -413,6 +413,11 @@ body::after {
     gap: 22px;
 }
 
+#main-interface {
+    height: 100%;
+    min-height: 0;
+}
+
 .landing-hero {
     border: 1px solid rgba(15, 159, 155, 0.18);
     background: linear-gradient(160deg, rgba(255, 255, 255, 0.9), rgba(255, 248, 238, 0.8));
@@ -509,6 +514,8 @@ body::after {
     grid-template-columns: minmax(0, 1.12fr) minmax(360px, 0.88fr);
     gap: 0;
     min-height: 0;
+    height: 100%;
+    align-items: stretch;
     overflow: hidden;
 }
 
@@ -517,7 +524,7 @@ body::after {
     flex-direction: column;
     gap: var(--hud-gap);
     padding: var(--hud-gap);
-    overflow-y: auto;
+    overflow-y: hidden;
     min-width: 0;
     min-height: 0;
 }
@@ -669,10 +676,15 @@ body::after {
     font-family: 'IBM Plex Mono', monospace;
     color: var(--ink-800);
     font-weight: 500;
+    overflow-wrap: anywhere;
+    word-break: break-word;
 }
 
 #action-panel {
     display: block;
+    max-height: 30vh;
+    overflow-y: auto;
+    padding-right: 2px;
 }
 
 #action-panel .stat-grid {
@@ -869,6 +881,7 @@ body::after {
     color: rgba(219, 233, 239, 0.9);
     position: relative;
     min-height: 0;
+    overscroll-behavior: contain;
 }
 
 .log-line {
@@ -972,6 +985,8 @@ body::after {
 .text-sm {
     font-size: 0.84rem;
     line-height: 1.6;
+    overflow-wrap: anywhere;
+    word-break: break-word;
 }
 
 ::-webkit-scrollbar {
@@ -1138,6 +1153,10 @@ body::after {
         padding: 20px;
         gap: 18px;
     }
+
+    .view-col {
+        overflow-y: auto;
+    }
 }
 
 @media (max-width: 700px) {

tests/test_graders.py

@@ -12,7 +12,8 @@ async def test_grade_easy_perfect_match():
     action = {"violation": "hate_speech"}
     ground_truth = {"violation": "hate_speech"}
     reward, feedback, rationale = await grade_action_async(action, ground_truth, "easy", {})
-    assert reward == 1.0
+    # Rewards are intentionally bounded by _smooth_reward to [0.05, 0.95].
+    assert reward == 0.95
     assert "Perfect" in feedback
     assert "Section 12" in rationale