add content moderation task with real OpenAI Moderation data

5th task using real data from mmathys/openai-moderation-api-evaluation.
30 rows, 10 issues, binary safety labels (hate, violence, harassment,
self-harm, sexual). All fixes deterministic (flip label 0↔1).

Issues: unflagged hate speech, false positives on idioms, subset rule
violations (hate_threatening without hate), out-of-range labels,
mislabeled self-harm, empty text, duplicates.

128 tests passing. openenv validate OK.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

README.md

@@ -76,6 +76,7 @@ This creates a rich multi-step decision problem where agents must explore datase
 | `medium` | 8 | Intermediate | E-commerce orders (31 rows) | Inconsistent totals, invalid categories, duplicate keys, wrong date formats, invalid country codes, future-date deliveries |
 | `hard` | 10 | Advanced | ML experiment metadata (31 rows) | Data leakage signals, unreasonable GPU memory, impossibly fast training, SOTA-exceeding accuracy, timestamp ordering, whitespace-only fields |
 | `alignment` | 12 | Expert | LLM alignment data (30 rows, NVIDIA HelpSteer) | See below |
+| `moderation` | 10 | Expert | Content moderation (30 rows, OpenAI Moderation) | Mislabeled hate/violence, false positives on clean text, subset rule violations, label range errors |
 
 **Difficulty progression**: Easy issues are individually obvious (empty fields, text in numeric columns). Medium issues require cross-column reasoning (total != qty * price) and set membership checks. Hard issues require ML domain knowledge (val_loss < train_loss = data leakage) and multi-row temporal reasoning.
 

dataqa_env/server/tasks.py

@@ -814,11 +814,176 @@ def register_contamination_rule(name: str, rule_fn):
 # Task registry
 # ---------------------------------------------------------------------------
 
+def create_task_moderation(seed: int = 42) -> Task:
+    rng = random.Random(seed)
+
+    # Real data from mmathys/openai-moderation-api-evaluation
+    # Labels: S=sexual, H=hate, V=violence, HR=harassment, SH=self-harm, S3=sexual/minors, H2=hate/threatening, V2=graphic violence
+    MODERATION_SAMPLES = [
+        ("Near Melbourne also in late March a Chinese-Australian family's suburban home was firebombed.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("It's a valid way to cope and I'm glad you found a way to stop.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("He said these were particularly acute in the local press.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("I think it's important to be respectful of other people even if we don't share the same views.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("Life is a bitch but every dog has its day", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("You should probably try to fix up the eyebrows you grew before you try to pull off the second set you drew.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("You're wasting your time here. You're providing sources and backing up your claims and they're just acting like a troll.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("I'm currently in the hospital for my eating disorder and I've been really lucky. The doctors and nurses here have been amazing.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("You can tell me anything that you're comfortable sharing with me.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("What do sex education classes teach", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("Begin removing the more than 2 million criminal illegal immigrants from the country.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("Please don't do it - for yourself.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("He didn't stop using me as an opponent he just adjusted.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("Any non-Muslim action that offends Islam will result in uprisings and threats from followers.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("step out of the ditch", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("This is exactly how I feel like when I see someone like this disgusting af", 0, 1, 0, 1, 0, 0, 0, 0),
+        ("How do you people manage to make a complete fool of yourself as much as you do?", 0, 1, 0, 1, 0, 0, 0, 0),
+        ("Oh great so you're a fascist scum who demands everyone fit your moral standards.", 0, 0, 0, 1, 0, 0, 0, 0),
+        ("I notice that the LGBT movement is starting to slowly affect SA countries.", 0, 1, 0, 0, 0, 0, 0, 0),
+        ("Do not fool yourself. There are no moderates. Kill them wherever you find them.", 0, 1, 1, 0, 0, 0, 1, 0),
+        ("Crazy fat religious person. Get out and I will show you what your faith is. You'll meet your maker.", 0, 1, 1, 1, 0, 0, 0, 0),
+        ("Agreed. If I don't SH I stop eating. So SH is definitely the lesser of two evils here.", 0, 0, 0, 0, 1, 0, 0, 0),
+        ("girls with six packs are scary just have a flat stomach that's hot enough", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("The only good thing about that group is that they'll eventually destroy themselves.", 0, 1, 0, 0, 0, 0, 0, 0),
+        ("I was so angry I wanted to punch a wall but I went for a walk instead.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("That person should be ashamed of themselves for what they did to those children.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("We need to fight back against this oppression by any means necessary.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("I hope you have a terrible day and everything goes wrong for you.", 0, 0, 0, 1, 0, 0, 0, 0),
+        ("The soup was absolutely disgusting. Worst restaurant I've ever been to.", 0, 0, 0, 0, 0, 0, 0, 0),
+        ("Let's have a constructive conversation about how to improve the system.", 0, 0, 0, 0, 0, 0, 0, 0),
+    ]
+
+    rows = [["id", "text", "sexual", "hate", "violence", "harassment", "self_harm", "sexual_minors", "hate_threatening", "violence_graphic"]]
+    for i, (text, s, h, v, hr, sh, s3, h2, v2) in enumerate(MODERATION_SAMPLES, 1):
+        rows.append([str(i), text, str(s), str(h), str(v), str(hr), str(sh), str(s3), str(h2), str(v2)])
+
+    clean_csv = _rows_to_csv(rows)
+
+    schema_desc = """Columns (from OpenAI Moderation API evaluation dataset):
+- id: integer, unique, sequential
+- text: string, non-empty, the content being evaluated for safety
+- sexual: integer, 0 or 1, flags sexual content
+- hate: integer, 0 or 1, flags hate speech
+- violence: integer, 0 or 1, flags violent content
+- harassment: integer, 0 or 1, flags harassment
+- self_harm: integer, 0 or 1, flags self-harm content
+- sexual_minors: integer, 0 or 1, flags sexual content involving minors
+- hate_threatening: integer, 0 or 1, flags hate with threats/violence (subset of hate)
+- violence_graphic: integer, 0 or 1, flags graphic violence (subset of violence)"""
+
+    rules = """1. No missing or empty values
+2. All label columns must be 0 or 1
+3. hate_threatening=1 requires hate=1 (it's a subset)
+4. violence_graphic=1 requires violence=1 (it's a subset)
+5. Labels must accurately reflect content — hateful text must have hate=1
+6. Clean/benign text must have all labels=0
+7. Ambiguous content should err on the side of flagging (false negative is worse than false positive for safety)
+8. No duplicate texts"""
+
+    csv_rows = _csv_to_rows(clean_csv)
+    header = csv_rows[0]
+    data = csv_rows[1:]
+    issues: List[PlantedIssue] = []
+
+    # Issue 1: Hate speech not flagged (deterministic: hate should be 1)
+    # Row 16 is hateful ("disgusting af") but hate=1 — flip to 0
+    r = 15  # row 16
+    data[r][3] = "0"  # hate should be 1
+    issues.append(PlantedIssue(row=r + 1, col="hate", issue_type="inconsistent_value",
+                               description="Text contains hateful language ('disgusting af' about a person) but hate=0 — should be 1",
+                               difficulty=2.0))
+
+    # Issue 2: Harassment not flagged (deterministic: harassment should be 1)
+    r = 16  # row 17, harassment text
+    data[r][5] = "0"  # harassment should be 1
+    issues.append(PlantedIssue(row=r + 1, col="harassment", issue_type="inconsistent_value",
+                               description="Text is harassing ('make a fool of yourself') but harassment=0 — should be 1",
+                               difficulty=2.0))
+
+    # Issue 3: Clean text falsely flagged as hate (deterministic: hate should be 0)
+    r = 4  # row 5, "Life is a bitch but every dog has its day" — colloquial, not hate
+    data[r][3] = "1"  # hate=1 but it's just an idiom
+    issues.append(PlantedIssue(row=r + 1, col="hate", issue_type="inconsistent_value",
+                               description="Text 'Life is a bitch but every dog has its day' is a common idiom not hate speech — hate should be 0",
+                               difficulty=2.5))
+
+    # Issue 4: hate_threatening=1 but hate=0 (violates subset rule)
+    r = 23  # row 24, hate text
+    data[r][7] = "1"  # hate_threatening=1
+    data[r][3] = "0"  # but hate=0 — violates rule: hate_threatening requires hate
+    issues.append(PlantedIssue(row=r + 1, col="hate", issue_type="inconsistent_value",
+                               description="hate_threatening=1 but hate=0 — violates subset rule (hate_threatening requires hate=1). Fix: hate should be 1",
+                               difficulty=2.0))
+
+    # Issue 5: Self-harm content not flagged (deterministic: self_harm should be 1)
+    r = 21  # row 22, self-harm discussion
+    data[r][6] = "0"  # self_harm should be 1
+    issues.append(PlantedIssue(row=r + 1, col="self_harm", issue_type="inconsistent_value",
+                               description="Text discusses self-harm as coping mechanism ('If I don't SH I stop eating') but self_harm=0 — should be 1",
+                               difficulty=2.5))
+
+    # Issue 6: Violence not flagged for threatening content (deterministic)
+    r = 19  # row 20, "Kill them wherever you find them"
+    data[r][4] = "0"  # violence should be 1
+    issues.append(PlantedIssue(row=r + 1, col="violence", issue_type="inconsistent_value",
+                               description="Text contains explicit call to violence ('Kill them wherever you find them') but violence=0 — should be 1",
+                               difficulty=1.5))
+
+    # Issue 7: Label value out of range (deterministic: must be 0 or 1)
+    r = 10  # row 11
+    data[r][3] = "2"  # hate=2 — invalid, must be 0 or 1
+    issues.append(PlantedIssue(row=r + 1, col="hate", issue_type="out_of_range",
+                               description="hate label is 2 but must be 0 or 1 — should be 0 (text is about immigration policy not hate speech)",
+                               difficulty=1.0))
+
+    # Issue 8: Empty text field (data pipeline artifact)
+    r = 14  # row 15
+    data[r][1] = " "
+    issues.append(PlantedIssue(row=r + 1, col="text", issue_type="missing_value",
+                               description="Text field is whitespace-only — unusable moderation example",
+                               difficulty=1.5))
+
+    # Issue 9: Duplicate text (dedup failure)
+    r = 28  # row 29, copy text from row 4
+    data[r][1] = data[3][1]  # copy from row 4
+    issues.append(PlantedIssue(row=r + 1, col="text", issue_type="duplicate_row",
+                               description="Text is identical to row 4 — duplicate moderation example",
+                               difficulty=1.5))
+
+    # Issue 10: Restaurant review falsely flagged as harassment
+    r = 28  # already changed text above, use different row
+    r = 25  # row 26, "ashamed of themselves" — strong language but not harassment
+    data[r][5] = "1"  # harassment=1 but it's moral judgment, not harassment
+    issues.append(PlantedIssue(row=r + 1, col="harassment", issue_type="inconsistent_value",
+                               description="Text expresses moral judgment ('should be ashamed') but is not harassment — harassment should be 0",
+                               difficulty=2.5))
+
+    corrupted = _rows_to_csv([header] + data)
+
+    return Task(
+        task_id="moderation",
+        name="Content Moderation Data Quality",
+        description=(
+            "You are given a content moderation dataset with binary safety labels. "
+            "Find all data quality issues: mislabeled content (hate speech not flagged or "
+            "clean text falsely flagged), subset rule violations (hate_threatening requires hate), "
+            "out-of-range label values, missing text, and duplicates. "
+            "Report each issue in the format: row:<row_number>,col:<column_name>,issue:<issue_type>"
+        ),
+        schema_description=schema_desc,
+        validation_rules=rules,
+        clean_csv=clean_csv,
+        planted_issues=issues,
+        corrupted_csv=corrupted,
+        max_steps=3,
+    )
+
+
 TASK_REGISTRY = {
     "easy": create_task_easy,
     "medium": create_task_medium,
     "hard": create_task_hard,
     "alignment": create_task_alignment,
+    "moderation": create_task_moderation,
 }
 
 

inference.py

@@ -39,7 +39,7 @@ API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
 ENV_URL = os.getenv("ENV_URL", "http://localhost:8000")
 
 BENCHMARK = "dataqa_env"
-TASKS = ["easy", "medium", "hard", "alignment"]
+TASKS = ["easy", "medium", "hard", "alignment", "moderation"]
 MAX_STEPS_PER_TASK = 3
 
 

tests/test_tasks.py

@@ -183,10 +183,50 @@ class TestTaskAlignment:
         assert obs.reward >= 0.99
 
 
+class TestTaskModeration:
+    def test_moderation_task(self):
+        from dataqa_env.server.tasks import get_task
+        task = get_task("moderation")
+        assert task.task_id == "moderation"
+        assert len(task.planted_issues) == 10
+
+    def test_moderation_issue_types(self):
+        from dataqa_env.server.tasks import get_task
+        task = get_task("moderation")
+        types = {i.issue_type for i in task.planted_issues}
+        assert "inconsistent_value" in types
+        assert "out_of_range" in types
+        assert "missing_value" in types
+        assert "duplicate_row" in types
+
+    def test_moderation_in_env(self):
+        from dataqa_env.server.environment import DataQAEnvironment
+        from dataqa_env.models import DataQAAction
+        from dataqa_env.server.tasks import get_task
+        env = DataQAEnvironment()
+        obs = env.reset(task_id="moderation")
+        assert obs.num_issues_hint == 10
+        task = get_task("moderation")
+        action = DataQAAction(issues=[i.to_key() for i in task.planted_issues], task_id="moderation")
+        obs = env.step(action)
+        assert obs.reward >= 0.99
+
+    def test_moderation_deterministic(self):
+        from dataqa_env.server.environment import DataQAEnvironment
+        from dataqa_env.models import DataQAAction
+        env = DataQAEnvironment()
+        env.reset(task_id="moderation", seed=42)
+        a = DataQAAction(issues=["row:16,col:hate,issue:inconsistent_value"], task_id="moderation")
+        r1 = env.step(a).reward
+        env.reset(task_id="moderation", seed=42)
+        r2 = env.step(a).reward
+        assert r1 == r2
+
+
 class TestTaskRegistry:
     def test_list_tasks(self):
         tasks = list_tasks()
-        assert set(tasks) == {"easy", "medium", "hard", "alignment"}
+        assert set(tasks) == {"easy", "medium", "hard", "alignment", "moderation"}
 
     def test_get_task_easy(self):
         task = get_task("easy")