File size: 14,400 Bytes
fcc38f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import re
from typing import Dict, List
import json



def parse_freethink(response: str, special_token_list=None, action_sep=',', max_actions=3) -> Dict:
    """
    Parse response in format: <think>...</think><answer>...</answer>
    
    Returns a dict with keys:
    - llm_raw_response: the original response
    - llm_response: the response with <think> and <answer> tags
    - think_content: the content inside <think> tag
    - action_content: the content inside <answer> tag
    - actions: a list of actions extracted from action_content
    - format_correct: whether the response strictly follows the expected format
    """
    response = response.replace("<image>","")
    #Pattern to check for content strictly in the format <think>...</think><answer>...</answer>
    strict_pattern = r'^\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*$'
    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)
    
    
    # Pattern to extract content from think and answer tags
    extraction_pattern = r'<think>(.*?)</think>\s*<answer>(.*?)</answer>'
    match = re.search(extraction_pattern, response, re.DOTALL)
    format_correct = strict_match is not None
    
    if not strict_match:
        think_content, action_content, actions = "", "", []
    else:
        think_content, action_content = match.group(1), match.group(2)
        if special_token_list is not None:
            for special_token in special_token_list: # remove all special tokens in responses to forbid confusion in training
                action_content = action_content.replace(special_token, "").strip()
                think_content = think_content.replace(special_token, "").strip()
        actions = [action.strip() for action in action_content.split(action_sep) if action.strip()]
        if len(actions) > max_actions:
            actions = actions[:max_actions] #Only the first MAX_ACTIONS actions are kept in the rollout.
            action_content = (" " + action_sep + " ").join(actions)

    llm_response = "<think>" + think_content.strip() + "</think>" + "<answer>" + action_content.strip() + "</answer>"
    return {
        "llm_raw_response": response,
        "llm_response": llm_response,
        "think_content": think_content,
        "action_content": action_content,
        "actions": actions,
        "format_correct": format_correct
    }

def parse_no_think(response: str, special_token_list=None, action_sep=',', max_actions=3) -> Dict:
    """
    Parse response in format: <answer>...</answer>
    
    Returns a dict with keys:
    - llm_raw_response: the original response
    - llm_response: the response with <answer> tag
    - think_content: empty string (no think content in this format)
    - action_content: the content inside <answer> tag
    - actions: a list of actions extracted from action_content
    - format_correct: whether the response strictly follows the expected format
    """
    response = response.replace("<image>","")
    # Pattern to check for content strictly in the format <answer>...</answer>
    strict_pattern = r'^\s*<answer>(.*?)</answer>\s*$'
    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)
    format_correct = strict_match is not None
    
    # Pattern to extract content from answer tag
    extraction_pattern = r'<answer>(.*?)</answer>'
    match = re.search(extraction_pattern, response, re.DOTALL)
    #format_correct = match is not None
    
    if not strict_match:
        think_content, action_content, actions = "", "", []
    else:
        action_content = match.group(1)
        think_content = ""  # No think content in this format
        if special_token_list is not None:
            for special_token in special_token_list:
                action_content = action_content.replace(special_token, "").strip()
        actions = [action.strip() for action in action_content.split(action_sep) if action.strip()]
        if len(actions) > max_actions:
            actions = actions[:max_actions]
            action_content = (" " + action_sep + " ").join(actions)

    llm_response = "<answer>" + action_content.strip() + "</answer>"
    return {
        "llm_raw_response": response,
        "llm_response": llm_response,
        "think_content": think_content,
        "action_content": action_content,
        "actions": actions,
        "format_correct": format_correct
    }

def parse_grounding(response: str, special_token_list=None, action_sep=',', max_actions=3) -> Dict:
    """
    Parse response in format: <think><observation>...</observation><reasoning>...</reasoning></think><answer>...</answer>
    
    Returns a dict with keys:
    - llm_raw_response: the original response
    - llm_response: the response with all tags
    - observation_content: the content inside <observation> tag
    - think_content: the entire content inside <think> tag
    - reasoning_content: the content inside <reasoning> tag
    - action_content: the content inside <answer> tag
    - actions: a list of actions extracted from action_content
    - format_correct: whether the response strictly follows the expected format
    """
    response = response.replace("<image>","")
    # Pattern to check for content strictly in the expected format
    strict_pattern = r'^\s*<think>\s*<observation>(.*?)</observation>\s*<reasoning>(.*?)</reasoning>\s*</think>\s*<answer>(.*?)</answer>\s*$'
    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)
    format_correct = strict_match is not None
    
    # Pattern to extract content from tags
    extraction_pattern = r'<think>\s*<observation>(.*?)</observation>\s*<reasoning>(.*?)</reasoning>\s*</think>\s*<answer>(.*?)</answer>'
    match = re.search(extraction_pattern, response, re.DOTALL)
    
    if not match:
        observation_content, reasoning_content, action_content, actions = "", "", "", []
        think_content = ""
    else:
        observation_content = match.group(1)
        reasoning_content = match.group(2)
        action_content = match.group(3)
        think_content = "<observation>" + observation_content + "</observation><reasoning>" + reasoning_content + "</reasoning>"
        
        if special_token_list is not None:
            for special_token in special_token_list:
                observation_content = observation_content.replace(special_token, "").strip()
                reasoning_content = reasoning_content.replace(special_token, "").strip()
                action_content = action_content.replace(special_token, "").strip()
                think_content = think_content.replace(special_token, "").strip()
                
        actions = [action.strip() for action in action_content.split(action_sep) if action.strip()]
        if len(actions) > max_actions:
            actions = actions[:max_actions]
            action_content = (" " + action_sep + " ").join(actions)
    
    # Reconstruct the cleaned llm_response
    llm_response = "<think>" + think_content.strip() + "</think>" + "<answer>" + action_content.strip() + "</answer>"
    
    return {
        "llm_raw_response": response,
        "llm_response": llm_response,
        "observation_content": observation_content,
        "think_content": think_content,
        "reasoning_content": reasoning_content,
        "action_content": action_content,
        "actions": actions,
        "format_correct": format_correct
    }

def parse_worldmodeling(response: str, special_token_list=None, action_sep=',', max_actions=3) -> Dict:
    """
    Parse response in format: <think><reasoning>...</reasoning><prediction>...</prediction></think><answer>...</answer>
    
    Returns a dict with keys:
    - llm_raw_response: the original response
    - llm_response: the response with all tags
    - think_content: the entire content inside <think> tag
    - reasoning_content: the content inside <reasoning> tag
    - prediction_content: the content inside <prediction> tag
    - action_content: the content inside <answer> tag
    - actions: a list of actions extracted from action_content
    - format_correct: whether the response strictly follows the expected format
    """
    response = response.replace("<image>","")
    # Pattern to check for content strictly in the expected format
    strict_pattern = r'^\s*<think>\s*<reasoning>(.*?)</reasoning>\s*<prediction>(.*?)</prediction>\s*</think>\s*<answer>(.*?)</answer>\s*$'
    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)
    format_correct = strict_match is not None
    
    # Pattern to extract content from tags
    extraction_pattern = r'<think>\s*<reasoning>(.*?)</reasoning>\s*<prediction>(.*?)</prediction>\s*</think>\s*<answer>(.*?)</answer>'
    match = re.search(extraction_pattern, response, re.DOTALL)
    
    if not match:
        reasoning_content, prediction_content, action_content, actions = "", "", "", []
        think_content = ""
    else:
        reasoning_content = match.group(1)
        prediction_content = match.group(2)
        action_content = match.group(3)
        think_content = "<reasoning>" + reasoning_content + "</reasoning><prediction>" + prediction_content + "</prediction>"
        
        if special_token_list is not None:
            for special_token in special_token_list:
                reasoning_content = reasoning_content.replace(special_token, "").strip()
                prediction_content = prediction_content.replace(special_token, "").strip()
                action_content = action_content.replace(special_token, "").strip()
                think_content = think_content.replace(special_token, "").strip()
                
        actions = [action.strip() for action in action_content.split(action_sep) if action.strip()]
        if len(actions) > max_actions:
            actions = actions[:max_actions]
            action_content = (" " + action_sep + " ").join(actions)
    
    # Reconstruct the cleaned llm_response
    llm_response = "<think>" + think_content.strip() + "</think>" + "<answer>" + action_content.strip() + "</answer>"
    
    return {
        "llm_raw_response": response,
        "llm_response": llm_response,
        "think_content": think_content,
        "reasoning_content": reasoning_content,
        "prediction_content": prediction_content,
        "action_content": action_content,
        "actions": actions,
        "format_correct": format_correct
    }

def parse_grounding_worldmodeling(response: str, special_token_list=None, action_sep=',', max_actions=3) -> Dict:
    """
    Parse response in format: <think><observation>...</observation><reasoning>...</reasoning><prediction>...</prediction></think><answer>...</answer>
    
    Returns a dict with keys:
    - llm_raw_response: the original response
    - llm_response: the response with all tags
    - observation_content: the content inside <observation> tag
    - reasoning_content: the content inside <reasoning> tag
    - prediction_content: the content inside <prediction> tag
    - think_content: the entire content inside <think> tag
    - action_content: the content inside <answer> tag
    - actions: a list of actions extracted from action_content
    - format_correct: whether the response strictly follows the expected format
    """
    response = response.replace("<image>","")
    # Pattern to check for content strictly in the expected format
    strict_pattern = r'^\s*<think>\s*<observation>(.*?)</observation>\s*<reasoning>(.*?)</reasoning>\s*<prediction>(.*?)</prediction>\s*</think>\s*<answer>(.*?)</answer>\s*$'
    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)
    format_correct = strict_match is not None
    
    # Pattern to extract content from tags
    extraction_pattern = r'<think>\s*<observation>(.*?)</observation>\s*<reasoning>(.*?)</reasoning>\s*<prediction>(.*?)</prediction>\s*</think>\s*<answer>(.*?)</answer>'
    match = re.search(extraction_pattern, response, re.DOTALL)
    
    if not match:
        observation_content, reasoning_content, prediction_content, action_content, actions = "", "", "", "", []
        think_content = ""
    else:
        observation_content = match.group(1)
        reasoning_content = match.group(2)
        prediction_content = match.group(3)
        action_content = match.group(4)
        think_content = "<observation>" + observation_content + "</observation><reasoning>" + reasoning_content + "</reasoning><prediction>" + prediction_content + "</prediction>"
        
        if special_token_list is not None:
            for special_token in special_token_list:
                observation_content = observation_content.replace(special_token, "").strip()
                reasoning_content = reasoning_content.replace(special_token, "").strip()
                prediction_content = prediction_content.replace(special_token, "").strip()
                action_content = action_content.replace(special_token, "").strip()
                think_content = think_content.replace(special_token, "").strip()
                
        actions = [action.strip() for action in action_content.split(action_sep) if action.strip()]
        if len(actions) > max_actions:
            actions = actions[:max_actions]
            action_content = (" " + action_sep + " ").join(actions)
    
    # Reconstruct the cleaned llm_response
    llm_response = "<think>" + think_content.strip() + "</think>" + "<answer>" + action_content.strip() + "</answer>"
    
    return {
        "llm_raw_response": response,
        "llm_response": llm_response,
        "observation_content": observation_content,
        "reasoning_content": reasoning_content,
        "prediction_content": prediction_content,
        "think_content": think_content,
        "action_content": action_content,
        "actions": actions,
        "format_correct": format_correct
    }
    
PARSE_FUNC_MAP = {
    "free_think": parse_freethink,
    "no_think": parse_no_think,
    "grounding": parse_grounding,
    "worldmodeling": parse_worldmodeling,
    "grounding_worldmodeling": parse_grounding_worldmodeling,
    "grounding_structured": parse_grounding,
    "worldmodeling_structured": parse_worldmodeling,
    "grounding_worldmodeling_structured": parse_grounding_worldmodeling,
    "grounding_symbolic": parse_grounding,
    "worldmodeling_symbolic": parse_worldmodeling,
    "grounding_worldmodeling_symbolic": parse_grounding_worldmodeling,
}