File size: 13,421 Bytes
a35c6f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json
from datasets import load_dataset

import verifiers as vf


def load_environment(
    num_train_examples=7000,
    num_eval_examples=1000,
    **kwargs
):
    """
    Environment for verifying complex JSON output from models.
    
    The task requires models to:
    1. Parse multi-question prompts
    2. Generate valid JSON responses
    3. Match the expected structure with correct keys and values
    
    Reward structure (multiplicative to prevent local minima):
    - If JSON fails to parse: reward = 0
    - Otherwise:
      * key_accuracy = (correct_keys) / (total_keys_in_response)
      * value_accuracy = (correct_values) / (total_values_in_response)
      * final_reward = key_accuracy * value_accuracy
    
    This penalizes both missing keys/values AND adding extra incorrect ones.
    """
    
    # Load dataset from HuggingFace
    dataset = load_dataset("Delta-Vector/Tauri-Complex-JSON-Formatting", split="train")
    
    # Sort by num_tasks (from metadata) to go from 2 to 14 in order
    # This helps identify where model collapses during training
    def extract_num_tasks(example):
        metadata = json.loads(example["metadata"])
        return metadata.get("num_tasks", 0)
    
    # Add num_tasks as a column for sorting
    dataset = dataset.map(lambda x: {"num_tasks": extract_num_tasks(x)})
    dataset = dataset.sort("num_tasks")
    
    # Map to expected format - keep verification_info as string to avoid schema issues
    def format_example(example):
        return {
            "question": example["prompt"],
            "info": {"verification_info": example["verification_info"]},  # Keep as dict with string
        }
    
    dataset = dataset.map(format_example, remove_columns=[col for col in dataset.column_names if col != "num_tasks"])
    
    # Split into train and eval (keeping the sorted order)
    train_dataset = dataset.select(range(num_train_examples))
    eval_dataset = dataset.select(range(num_train_examples, num_train_examples + num_eval_examples))
    
    # Custom extract function to parse JSON from code blocks or raw text
    def extract_json_from_completion(completion):
        """Extract JSON from completion, handling code blocks."""
        if not completion:
            return ""
        
        # Get the last message content
        if isinstance(completion, list) and len(completion) > 0:
            content = completion[-1].get("content", "")
        else:
            content = str(completion)
        
        # Try to extract from code blocks first (```json ... ``` or ``` ... ```)
        import re
        code_block_pattern = r"```(?:json)?\s*\n(.*?)\n```"
        matches = re.findall(code_block_pattern, content, re.DOTALL)
        if matches:
            return matches[-1].strip()  # Return last code block
        
        # Otherwise return the content as-is
        return content.strip()
    
    # Use simple Parser with custom extract function
    parser = vf.Parser(extract_fn=extract_json_from_completion)
    
    def multiplicative_reward(completion, info, **kwargs) -> float:
        """
        Multiplicative reward: key_accuracy * value_accuracy.
        
        Returns 0 if JSON fails to parse.
        Otherwise:
        - key_accuracy = (correct_keys) / (total_keys_in_response)
        - value_accuracy = (correct_values) / (total_values_in_response)
        - final_reward = key_accuracy * value_accuracy
        
        This penalizes both missing correct items AND adding extra incorrect ones.
        """
        try:
            response = parser.parse_answer(completion) or ""
            response = response.strip()
            
            # Check: Valid JSON format
            if not response:
                return 0.0
            
            try:
                parsed_response = json.loads(response)
            except (json.JSONDecodeError, ValueError):
                return 0.0
            
            # Must be a dict
            if not isinstance(parsed_response, dict):
                return 0.0
            
            # Parse ground truth from info
            verification_info = json.loads(info["verification_info"])
            ground_truth = verification_info["ground_truth"]
            
            # Get all keys recursively with their full paths
            def get_all_keys(d, prefix=""):
                keys = set()
                if isinstance(d, dict):
                    for k, v in d.items():
                        full_key = f"{prefix}.{k}" if prefix else k
                        keys.add(full_key)
                        keys.update(get_all_keys(v, full_key))
                return keys
            
            # Get all values recursively
            def get_all_values(d):
                values = []
                if isinstance(d, dict):
                    for v in d.values():
                        if isinstance(v, dict):
                            values.extend(get_all_values(v))
                        elif isinstance(v, list):
                            values.extend(get_all_values({"_": item} for item in v))
                        else:
                            values.append(v)
                return values
            
            ground_truth_keys = get_all_keys(ground_truth)
            response_keys = get_all_keys(parsed_response)
            
            # Calculate key accuracy
            if len(response_keys) == 0:
                key_accuracy = 0.0
            else:
                correct_keys = len(ground_truth_keys & response_keys)  # Intersection
                key_accuracy = correct_keys / len(response_keys)
            
            # Calculate value accuracy by checking each value at correct key paths
            def get_value_at_path(d, path):
                """Get value at a specific key path like 'a.b.c'"""
                keys = path.split('.')
                current = d
                try:
                    for key in keys:
                        current = current[key]
                    return current
                except (KeyError, TypeError):
                    return None
            
            # Helper function to compare values with numeric type tolerance
            def values_equal(a, b):
                """Compare values with numeric type tolerance (25 == 25.0)"""
                # Handle numeric comparison (int vs float)
                if isinstance(a, (int, float)) and isinstance(b, (int, float)):
                    return a == b  # Python handles int/float equality correctly
                # For everything else, use strict equality
                return a == b
            
            total_values_checked = len(response_keys)
            
            if total_values_checked == 0:
                value_accuracy = 0.0
            else:
                correct_values = 0
                for key_path in response_keys:
                    response_val = get_value_at_path(parsed_response, key_path)
                    ground_truth_val = get_value_at_path(ground_truth, key_path)
                    
                    # If key exists in ground truth and values match
                    if ground_truth_val is not None and values_equal(response_val, ground_truth_val):
                        correct_values += 1
                
                value_accuracy = correct_values / total_values_checked
            
            # Multiply together
            final_reward = key_accuracy * value_accuracy
            return final_reward
                
        except (AttributeError, TypeError, KeyError):
            return 0.0
    
    def format_reward(completion, **kwargs) -> float:
        """
        Reward for valid JSON formatting.
        Returns 0.33 for valid JSON dict, 0 for invalid.
        """
        try:
            response = parser.parse_answer(completion) or ""
            response = response.strip()
            
            # Check if response is not empty
            if not response:
                return 0.0
            
            # Try to parse as JSON
            parsed = json.loads(response)
            
            # Must be a dict (since ground truth is always a dict)
            if not isinstance(parsed, dict):
                return 0.0
            
            return 0.33
        except (json.JSONDecodeError, ValueError, TypeError):
            return 0.0
    
    def keys_match_reward(completion, info, **kwargs) -> float:
        """
        Metric: key accuracy (correct_keys / total_keys_in_response).
        Returns the same key_accuracy used in multiplicative_reward.
        """
        try:
            response = parser.parse_answer(completion) or ""
            response = response.strip()
            
            if not response:
                return 0.0
                
            parsed_response = json.loads(response)
            
            if not isinstance(parsed_response, dict):
                return 0.0
            
            # Parse ground truth from info
            verification_info = json.loads(info["verification_info"])
            ground_truth = verification_info["ground_truth"]
            
            # Get all keys from ground truth (recursively)
            def get_all_keys(d, prefix=""):
                keys = set()
                if isinstance(d, dict):
                    for k, v in d.items():
                        full_key = f"{prefix}.{k}" if prefix else k
                        keys.add(full_key)
                        keys.update(get_all_keys(v, full_key))
                return keys
            
            ground_truth_keys = get_all_keys(ground_truth)
            response_keys = get_all_keys(parsed_response)
            
            if len(response_keys) == 0:
                return 0.0
            
            correct_keys = len(ground_truth_keys & response_keys)
            return correct_keys / len(response_keys)
                
        except (json.JSONDecodeError, ValueError, AttributeError, TypeError):
            return 0.0
    
    def values_match_reward(completion, info, **kwargs) -> float:
        """
        Metric: value accuracy (correct_values / total_values_in_response).
        Returns the same value_accuracy used in multiplicative_reward.
        """
        try:
            response = parser.parse_answer(completion) or ""
            response = response.strip()
            
            if not response:
                return 0.0
                
            parsed_response = json.loads(response)
            
            if not isinstance(parsed_response, dict):
                return 0.0
            
            # Parse ground truth from info
            verification_info = json.loads(info["verification_info"])
            ground_truth = verification_info["ground_truth"]
            
            # Helper function to compare values with numeric type tolerance
            def values_equal(a, b):
                if isinstance(a, (int, float)) and isinstance(b, (int, float)):
                    return a == b
                return a == b
            
            # Get all keys recursively
            def get_all_keys(d, prefix=""):
                keys = set()
                if isinstance(d, dict):
                    for k, v in d.items():
                        full_key = f"{prefix}.{k}" if prefix else k
                        keys.add(full_key)
                        keys.update(get_all_keys(v, full_key))
                return keys
            
            def get_value_at_path(d, path):
                keys = path.split('.')
                current = d
                try:
                    for key in keys:
                        current = current[key]
                    return current
                except (KeyError, TypeError):
                    return None
            
            response_keys = get_all_keys(parsed_response)
            
            if len(response_keys) == 0:
                return 0.0
            
            correct_values = 0
            for key_path in response_keys:
                response_val = get_value_at_path(parsed_response, key_path)
                ground_truth_val = get_value_at_path(ground_truth, key_path)
                
                if ground_truth_val is not None and values_equal(response_val, ground_truth_val):
                    correct_values += 1
            
            return correct_values / len(response_keys)
                
        except (json.JSONDecodeError, ValueError, AttributeError, TypeError):
            return 0.0
    
    # Create rubric with multiplicative reward
    # Keep individual functions for debugging/metrics but use multiplicative for training
    rubric = vf.Rubric(
        parser=parser,
        funcs=[
            multiplicative_reward,   # Main reward - key_acc * value_acc
            format_reward,           # Metric only (weight 0)
            keys_match_reward,       # Metric only (weight 0)
            values_match_reward,     # Metric only (weight 0)
        ],
        weights=[1.0, 0.0, 0.0, 0.0]  # Only multiplicative_reward counts
    )
    
    # Return SingleTurnEnv since this is a one-shot task
    # No system prompt - let the dataset prompt speak for itself
    vf_env = vf.SingleTurnEnv(
        dataset=train_dataset,
        eval_dataset=eval_dataset,
        parser=parser,
        rubric=rubric,
    )
    
    return vf_env