File size: 19,738 Bytes
8ced877
0c216ef
f5583f9
0c216ef
 
8ced877
 
0bb6f99
0c216ef
 
8ced877
 
0c216ef
 
6c1b2ac
f5583f9
6c1b2ac
f5583f9
6c1b2ac
0c216ef
6c1b2ac
0c216ef
6c1b2ac
0c216ef
6c1b2ac
 
 
 
 
 
 
 
 
0c216ef
6c1b2ac
0c216ef
6c1b2ac
0c216ef
6c1b2ac
0c216ef
6c1b2ac
 
f5583f9
 
 
 
 
6c1b2ac
f5583f9
6c1b2ac
0c216ef
6c1b2ac
0c216ef
6c1b2ac
0c216ef
6c1b2ac
 
 
 
0c216ef
6c1b2ac
0c216ef
6c1b2ac
0c216ef
6c1b2ac
 
 
0c216ef
6c1b2ac
 
 
 
 
 
 
0c216ef
6c1b2ac
f5583f9
6c1b2ac
f5583f9
6c1b2ac
f5583f9
6c1b2ac
 
 
 
 
 
 
 
 
f5583f9
6c1b2ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f5583f9
346b0b1
 
6c1b2ac
f5583f9
6c1b2ac
 
346b0b1
6c1b2ac
 
 
f5583f9
346b0b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f5583f9
346b0b1
 
 
f5583f9
346b0b1
f5583f9
346b0b1
f5583f9
346b0b1
 
 
 
 
 
 
 
 
 
 
f5583f9
6c1b2ac
 
 
 
 
 
 
f5583f9
346b0b1
 
 
 
 
 
 
 
 
f5583f9
346b0b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0c216ef
 
346b0b1
0c216ef
346b0b1
 
0c216ef
 
346b0b1
f5583f9
 
 
0c216ef
 
 
346b0b1
f5583f9
346b0b1
 
 
 
 
 
6c1b2ac
f5583f9
 
 
 
 
346b0b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0c216ef
 
 
 
346b0b1
 
 
0c216ef
346b0b1
6c1b2ac
346b0b1
 
 
 
 
 
 
 
 
 
 
 
0c216ef
346b0b1
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
---
title: DataQA Environment Server
emoji: "\U0001F50D"
colorFrom: blue
colorTo: gray
sdk: docker
pinned: false
app_port: 8000
tags:
  - openenv
---

# DataQA Environment

**A two-phase OpenEnv RL environment for Data Quality Assurance** β€” an LLM agent inspects corrupted datasets, identifies all planted quality issues, and proposes data repairs.

## Why DataQA? The Moat

### 1. Solves a Real, High-Frequency Problem

Every ML team burns hours on data quality β€” missing values, type mismatches, logical inconsistencies, subtle statistical anomalies β€” before data enters training pipelines or production databases. DataQA turns this universal pain point into a graded RL environment. Unlike synthetic toy problems, **these are the exact data bugs that corrupt production ML models.**

### 2. Seven Diverse Domains, One Unified Interface

| Task | Domain | Issues | What Makes It Hard |
|------|--------|--------|--------------------|
| `easy` | HR / Employee data | 6 | Missing values, typos, format errors |
| `medium` | E-commerce orders | 8 | Cross-column math (`total != qty * price`), OCR errors |
| `hard` | ML experiment metadata | 10 | Data leakage detection, impossible GPU specs, SOTA violations |
| `alignment` | LLM fine-tuning data (NVIDIA HelpSteer) | 12 | Hallucinated citations, self-contradictions, toxic content scored as helpful |
| `coding` | Code instruction-response pairs | 10 | Logic bugs in "correct" code, `eval()` injection, language mismatches |
| `toolcalling` | Function-calling schemas | 10 | Hallucinated parameters, missing required args, name mismatches |
| `moderation` | Content moderation labels | 10 | Mislabeled hate speech, false positives on clean text |

**66 total planted issues** spanning tabular data, free-text, code, JSON schemas, and safety labels. No other OpenEnv submission covers this breadth with a single coherent reward function.

### 3. Two-Phase Reward β€” Identify Then Fix

Most data QA environments only ask "is there a bug?" DataQA goes further:

- **Phase 1 (Identify):** Find all issues β€” graded by difficulty-weighted F1
- **Phase 2 (Fix):** Propose the correct value β€” graded against the clean original with tiered scoring (exact match = 1.0, valid fix = 0.8, partial = 0.4, right cell wrong value = 0.1)

```
combined_reward = 0.6 * identify_score + 0.4 * fix_score
```

This creates a richer learning signal than binary classification. An agent that finds 8/10 issues and fixes 5 of them correctly gets meaningful partial credit β€” perfect for GRPO/RLHF training.

### 4. Difficulty-Weighted Scoring Rewards Deeper Reasoning

Each planted issue has a difficulty weight (1.0-3.0). Finding a hallucinated citation (3.0) earns triple the reward of finding an empty field (1.0). This incentivizes agents to develop genuine reasoning capabilities rather than pattern-matching surface-level errors.

### 5. Multi-Step Feedback Loop

Agents get 3 attempts per task with detailed per-step feedback:
- Which issues were correct (true positives) vs wrong (false positives)
- Which issues were missed (false negatives) with difficulty hints
- Fix quality scores with reasons

This enables the agent to **learn from its mistakes within a single episode** β€” a natural curriculum.

### 6. Fully Extensible

```python
# Add your own contamination rules
register_contamination_rule("swap_digits", my_swap_fn)

# Create tasks from any CSV
task = create_task_from_config(
    task_id="custom", clean_csv="...",
    contaminations=[{"rule": "missing_value", "row": 0, "col": 1}]
)
register_task("custom", lambda seed: task)
```

New domains can be added in minutes. The contamination engine is domain-agnostic.

---

## Demo: Agent Trajectory

```
HARD TASK β€” ML experiment metadata
  Step 1: Found 5/10, missed hard issues    β†’ Reward: 0.69
  Step 2: Found 10/10 + 5 fixes proposed   β†’ Reward: 0.77
  Issues requiring ML knowledge:
    β€’ val_loss < train_loss (data leakage signal)
    β€’ resnet18 using 42.5GB GPU (impossible for 11M params)
    β€’ 350 epochs on ImageNet in 30 min (impossibly fast)
    β€’ wav2vec2 at 98.5% accuracy (exceeds SOTA)

ALIGNMENT TASK β€” NVIDIA HelpSteer data
  Step 1: Found 7/12, missed subtle issues  β†’ Reward: 0.58
  Step 2: Found 12/12 + 3 fixes proposed   β†’ Reward: 0.72
  Issues requiring deep reasoning:
    β€’ Cerasus vs Prunus serrulata (wrong taxonomic name)
    β€’ $400.3M at Sotheby's vs $450.3M at Christie's (close but wrong)
    β€’ Fake Nature paper by "Dr. Sarah Chen" (hallucinated citation)
    β€’ Gender-biased advice rated helpfulness=4 (toxic content with inflated scores)

CODING TASK β€” Code instruction-response pairs
  Issues requiring code understanding:
    β€’ Binary search off-by-one (lo=mid causes infinite loop) marked correct
    β€’ eval(uid) in Flask route β€” code injection vulnerability
    β€’ JavaScript response for a Python-labeled task
    β€’ Duplicate "merge sort" instruction across rows
```

> The interactive replay UI with color-coded dataset visualization is available on the HF Space.

## Environment API

| Endpoint | Method | Description |
|----------|--------|-------------|
| `/reset` | POST | Start a new episode with a corrupted dataset |
| `/step` | POST | Submit identified issues + proposed fixes |
| `/state` | GET | Get current episode state |
| `/health` | GET | Health check |

## Tasks

**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). Expert tasks (alignment, coding, toolcalling, moderation) require domain expertise, semantic reasoning, and cross-row comparison.

### Alignment Task: LLM Training Data Quality (Expert)

Built on **real data from [NVIDIA HelpSteer](https://huggingface.co/datasets/nvidia/HelpSteer)** β€” 30 human-annotated prompt-response pairs with quality scores (helpfulness, correctness, coherence, complexity, verbosity on 0-4 scale).

This task targets a critical real-world problem: **catching quality issues in LLM fine-tuning data before it corrupts model training**. The 12 planted issues represent failure modes actually seen in production data pipelines:

| Issue | Difficulty | Why It's Hard |
|---|---|---|
| Subtle factual error (*Cerasus* vs *Prunus serrulata*) | 3.0 | Old taxonomic synonym β€” sounds plausible, requires domain knowledge |
| Plausible wrong numbers ($400.3M at Sotheby's vs $450.3M at Christie's) | 3.0 | Right painting, wrong price by $50M and wrong auction house |
| Self-contradictory reasoning ("does NOT learn via backprop" then describes backprop) | 3.0 | Response negates its own conclusion β€” trains confused models |
| Hallucinated citation (fake Nature paper by fake Dr. Sarah Chen) | 3.0 | Fabricated study with specific fake statistics β€” most dangerous for training |
| Harmful coding advice ("use bare except everywhere") with high quality scores | 3.0 | Teaches dangerous practices if used for fine-tuning |
| Toxic/biased response scored as helpful | 3.0 | Gender-biased stereotypes with helpfulness=4 β€” poisons alignment training |
| Leaked system prompt (`[SYSTEM] You are a helpful AI...`) in response | 2.5 | Data pipeline failed to strip prompt template |
| Semantic near-duplicate prompt (rephrased, not exact copy) | 2.5 | Requires semantic similarity detection, not just string matching |
| Truncated response (cut mid-sentence) | 2.5 | `max_length` truncation without sentence boundary detection |
| Response in French for English prompt | 2.0 | Language contamination from multilingual training data |
| Response plagiarized from another row | 2.0 | Data pipeline shuffling/dedup failure |
| Whitespace-only prompt | 2.0 | Empty training example from pipeline artifact |

### Coding Task: Code Quality (Expert)

20-row dataset of code instruction-response pairs (Python algorithms, data structures, web, design patterns). 10 planted issues:

- Syntax errors in "correct" code (unbalanced parens)
- Logic bugs marked `is_correct=true` (binary search off-by-one infinite loop)
- Security vulnerabilities (`eval()` on user input) marked correct
- Language mismatches (JavaScript response labeled Python)
- Truncated code, difficulty label mismatches, duplicate instructions, wrong categories, missing test cases

### Tool-Calling Task: Function Schema Quality (Expert)

20-row dataset of function definitions with parameter schemas, example calls, and outputs. 10 planted issues:

- Function name mismatch between definition and example call
- Missing required parameters in example call
- Hallucinated parameters not in schema
- Type mismatches (string "high" for integer quality parameter)
- Invalid JSON, duplicate function names, misleading descriptions, wrong categories

### Moderation Task: Content Label Quality (Expert)

30-row dataset modeled on content moderation pipelines. 10 planted issues:

- Mislabeled hate speech and violence (unflagged toxic content)
- False positives on clean text (idioms flagged as hate)
- Subset rule violations (`hate_threatening` without `hate` flag)
- Out-of-range label values

## Two-Phase Action Space

### Phase 1: Identify Issues

Submit issues in format: `row:<row_number>,col:<column_name>,issue:<issue_type>`

- `row_number`: 1-indexed data row position (after header)
- `column_name`: Exact column header name, lowercase
- `issue_type`: One of the supported types below

### Phase 2: Propose Fixes

Submit fixes in format: `row:<row_number>,col:<column_name>,fix:<corrected_value>`

The agent proposes the **correct value** that should replace the corrupted data. Fixes are graded against the original clean dataset.

Both phases can be submitted in the same step or across multiple steps.

**Supported Issue Types:**

| Type | Description | Example |
|------|-------------|---------|
| `missing_value` | Null, empty, or whitespace-only | Empty name field |
| `wrong_type` | Value doesn't match expected type | Salary as "seventy-five thousand" |
| `duplicate_row` | Exact duplicate or duplicate key | Two rows with same employee_id |
| `out_of_range` | Value outside valid range | Salary of 5000 when min is 50000 |
| `format_violation` | Wrong format or invalid enum | Date as DD/MM/YYYY instead of YYYY-MM-DD |
| `inconsistent_value` | Computed field mismatch, logical inconsistency | total != qty * price |
| `statistical_outlier` | Unreasonable value given context | resnet18 using 42.5GB GPU |
| `referential_integrity` | Foreign key violation | (available for custom tasks) |

## Observation Space

| Field | Type | Description |
|-------|------|-------------|
| `dataset_csv` | str | The corrupted dataset in CSV format |
| `schema_description` | str | Column types, ranges, and constraints |
| `validation_rules` | str | Business rules the data must satisfy |
| `task_description` | str | Task context and instructions |
| `feedback` | str | Per-step results: TP/FP/FN, precision/recall, fix scores |
| `num_issues_hint` | int | Exact count of planted issues |
| `max_steps` | int | Maximum attempts allowed |
| `done` | bool | Whether episode has terminated |
| `reward` | float | Best combined reward so far (strict 0-1 range) |

**Observation Metadata** (per step):
- Identify: `identify_f1`, `identify_score`, `precision`, `recall`, `tp`, `fp`, `fn`
- Fix: `fix_score`, `fixes_correct`, `fixes_partial`, `fixes_wrong`, `fixes_attempted`
- Combined: `combined_reward`, `difficulty_found`, `difficulty_missed`

## Reward Function

### Combined Reward

```
combined_reward = 0.6 * identify_score + 0.4 * fix_score
```

If no fixes are submitted, `combined_reward = identify_score` (no penalty β€” backward compatible).

### Identify Score (Difficulty-Weighted F1)

Each planted issue has a **difficulty weight** (1.0-3.0):

| Weight | Category | Examples |
|--------|----------|----------|
| 1.0 | Easy | Missing values, obvious out-of-range, wrong type |
| 1.5-2.0 | Medium | Duplicate keys, format violations, cross-column checks |
| 2.5-3.0 | Hard | Data leakage, statistical outliers, hallucinated citations |

- **Weighted Recall** = (difficulty of found issues) / (total difficulty)
- **Weighted Precision** = penalizes false positives proportional to average difficulty
- **Weighted F1** = harmonic mean

### Fix Score (Tiered Grading by Issue Type)

Each proposed fix is graded with tiered scoring that gives partial credit for reasonable attempts:

| Fix Quality | Score | Description |
|-------------|-------|-------------|
| Exact match | 1.0 | Case-insensitive, whitespace-stripped match with clean value |
| Valid fix | 0.8 | Right type/range, addresses the issue (e.g., any non-empty value for missing field) |
| Partially valid | 0.4 | Reasonable attempt, right direction (e.g., numeric in right ballpark) |
| Right cell, wrong value | 0.1 | Targets correct cell but fix doesn't address the issue |
| Non-issue cell | 0.0 | Fix targets a cell with no issue |

Fix score = (sum of best fix score per issue x difficulty weight) / (total difficulty weight)

### Reward Properties

| Property | Detail |
|----------|--------|
| Range | Strict (0, 1) β€” 0.001 minimum, 0.999 maximum |
| Partial credit | Yes β€” per-issue, difficulty-weighted |
| Monotonic | Best score across all steps is final reward |
| Penalizes guessing | False positives reduce precision, fixing non-issues scores 0 |
| Multi-step improvement | Detailed feedback enables learning across attempts |

### Episode Boundaries

- Each task allows up to 3 steps (attempts)
- Episode ends when F1 >= 0.999 (perfect identification) or max steps reached
- Agent receives detailed feedback after each step to improve on next attempt

## Extensibility

### Custom Contamination Rules

```python
from dataqa_env import register_contamination_rule
from dataqa_env.server.tasks import PlantedIssue

def swap_digits(rows, header, col_idx, row_idx, rng):
    val = rows[row_idx][col_idx]
    corrupted = val[::-1]
    issue = PlantedIssue(
        row=row_idx + 1, col=header[col_idx],
        issue_type="format_violation",
        description=f"Digits swapped in {header[col_idx]}",
        difficulty=2.0,
    )
    return corrupted, issue

register_contamination_rule("swap_digits", swap_digits)
```

### Custom Tasks from Config

```python
from dataqa_env import create_task_from_config, register_task

task = create_task_from_config(
    task_id="custom",
    name="Custom Validation",
    description="Find quality issues in this dataset.",
    schema_description="id: int, name: str, score: int (0-100)",
    validation_rules="No missing values. Scores must be 0-100.",
    clean_csv="id,name,score\n1,Alice,95\n2,Bob,87\n3,Carol,92",
    contaminations=[
        {"rule": "missing_value", "row": 0, "col": 1, "difficulty": 1.0},
        {"rule": "negative_value", "row": 2, "col": 2, "difficulty": 1.5},
    ],
)
register_task("custom", lambda seed: task)
```

### Built-in Contamination Rules

| Rule | Effect | Default Difficulty |
|------|--------|--------------------|
| `missing_value` | Sets field to empty string | 1.0 |
| `whitespace_value` | Sets field to single space | 2.5 |
| `wrong_type_text` | Replaces with random text | 1.0 |
| `negative_value` | Negates numeric value | 1.0 |

## Setup & Quick Start

```bash
# Install
pip install -e .

# Run server locally
uvicorn dataqa_env.server.app:app --host 0.0.0.0 --port 8000

# Run inference (set your API credentials)
API_BASE_URL=https://router.huggingface.co/v1 \
MODEL_NAME=Qwen/Qwen2.5-72B-Instruct \
HF_TOKEN=your-token \
python inference.py
```

## Docker

```bash
docker build -t dataqa-env .
docker run -p 8000:8000 dataqa-env
```

## Testing

```bash
pip install -e ".[dev]"
pytest tests/ -v
```

128 tests covering:
- Task creation, corruption, and difficulty weights for all 7 tasks
- Issue key and fix parsing (standard, lenient, edge cases)
- F1, weighted reward, and fix quality computation
- Full environment lifecycle (identify-only and identify+fix)
- Combined reward calculation and weight verification
- Inference script parsing and prompt building
- Structured log format ([START], [STEP], [END])
- Score bounds (strict 0-1), best-score monotonicity
- Extensibility API (custom rules, custom tasks)
- Moderation task determinism and label consistency

## Validation

```bash
# OpenEnv spec validation
openenv validate .

# Pre-submission validation (requires HF Space URL)
./prevalidation_script.sh https://your-space.hf.space
```

## Environment Variables

| Variable | Description | Default |
|----------|-------------|---------|
| `API_BASE_URL` | LLM API endpoint | `https://router.huggingface.co/v1` |
| `MODEL_NAME` | Model identifier | `Qwen/Qwen2.5-72B-Instruct` |
| `HF_TOKEN` | HuggingFace token / API key | - |
| `ENV_URL` | Environment server URL | `http://localhost:8000` |

## Architecture

```
dataqa_env/
β”œβ”€β”€ __init__.py            # Public API + extensibility exports
β”œβ”€β”€ models.py              # Pydantic: DataQAAction (issues + fixes), DataQAObservation, DataQAState
β”œβ”€β”€ client.py              # EnvClient for WebSocket connections
β”œβ”€β”€ server/
β”‚   β”œβ”€β”€ environment.py     # Two-phase DataQAEnvironment (identify + fix + combined reward)
β”‚   β”œβ”€β”€ tasks.py           # 7 task definitions + contamination rules + extensibility API
β”‚   β”œβ”€β”€ gradio_ui.py       # Interactive web UI with agent trajectory replay
β”‚   β”œβ”€β”€ app.py             # FastAPI server (via openenv-core create_app)
β”‚   └── Dockerfile
tests/
β”œβ”€β”€ test_tasks.py          # Task creation, corruption, difficulty weights (all 7 tasks)
β”œβ”€β”€ test_environment.py    # Identify scoring, fix grading, combined reward, lifecycle
β”œβ”€β”€ test_inference.py      # LLM response parsing, fix parsing, prompt building, log format
└── test_extensibility.py  # Custom rules, custom tasks, registration API
inference.py               # Two-phase baseline agent (identify then fix)
openenv.yaml               # OpenEnv/HF Spaces spec
pyproject.toml             # Package metadata and dependencies
Dockerfile                 # Production container
```

### Key Modules

**`dataqa_env/server/tasks.py`** β€” The core of the environment. Each task function (`create_task_easy`, `create_task_coding`, etc.) builds a clean CSV dataset, injects corruptions as `PlantedIssue` objects with row/col/type/difficulty, and returns a `Task` dataclass. The `TASK_REGISTRY` dict maps task IDs to factory functions. The extensibility API (`register_task`, `register_contamination_rule`, `create_task_from_config`) allows users to add domains without modifying source.

**`dataqa_env/server/environment.py`** β€” The `DataQAEnvironment` class inherits from OpenEnv's `Environment` base. `reset()` loads a task by ID and returns the corrupted CSV + schema. `step()` parses issue keys and fix proposals from the action, computes difficulty-weighted F1 for identification, grades fixes with tiered scoring by issue type, and returns combined reward with detailed feedback. Handles HTTP statelessness via auto-reset from `action.task_id`.

**`dataqa_env/models.py`** β€” Pydantic models for the OpenEnv interface. `DataQAAction` carries `issues: List[str]`, `fixes: List[str]`, and `task_id: str`. `DataQAObservation` carries the CSV, schema, rules, feedback, and scoring metadata. `DataQAState` tracks episode progress.

**`inference.py`** β€” Baseline LLM agent using OpenAI-compatible API. Runs all 7 tasks sequentially with 3 steps each. Lenient regex parsing handles case variations and delimiter differences in LLM output. Structured logging in `[START]/[STEP]/[END]` format for evaluation.