Robo-CSK-Benchmark/main.py at main · ag-sc/Robo-CSK-Benchmark · GitHub

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
import argparse
import os
import json
import dspy

from meta_reasoning.prompting import meta_reason_prompter_binary, meta_reason_prompter_multi_choice
from procedural_knowledge.prompting import cooking_procedures_prompter
from table_setting.prompting import table_setting_prompter
from tidy_up.prompting import tidy_up_prompter_open, tidy_up_prompter_multi_choice
from tool_usage.prompting import tool_use_prompter
from utils import GemmaPrompter, OpenAIPrompter, LlamaPrompter

it_selfcon = 3
it_selfref = 5
n_ex_sgicl = 8
n_ex_contr = 4

def main():
    parser = argparse.ArgumentParser(
        description='RoboCSKBench - Please choose the model to evaluate and the tasks to evaluate on')
    parser.add_argument("--models", type=str, choices=["gpt", "llama", "gemma", "dspy"], nargs="+",
                        help="Choose one or more models to run")
    parser.add_argument("--tasks", type=str, choices=["ts", "mr_b", "mr_mc", "to_us", "ti_up_o", "ti_up_mc", "proc_b", "proc_mc"],
                        nargs="+", help="Choose one or more tasks to run")
    parser.add_argument('--techs' ,type=str, choices=['base', 'rar', 'meta', 'contr', 'selfcon', 'selfref', 'sgicl', 'stepback'],
                        nargs="+", help='Choose one or more techniques to prompt the model with')
    parser.add_argument('--mode', type=str, choices=['scratch', 'role', 'eval', 'show', 'base'],
                        help='Select between optimizing from scratch or role based prompt, evaluating current saved program and showing current saved prompts')
    parser.add_argument('--lower', type=int, default=1, help='Lower bound of custom evaluation window')
    parser.add_argument('--upper', type=int, default=None, help='Upper bound of custom evaluation window (exclusive)')
    parser.add_argument("--new_tok", type=int, default=1000, help="Maximum number of new tokens for the model output")
    parser.add_argument("--num_q", type=int, default=None, help="Number of questions to be asked in each task. Testing only")
    args = parser.parse_args()

    num_runs = args.num_q
    lower = args.lower
    upper = args.upper

    # For compatibility with pd.read()
    if upper == None:
        bound = None
    else:
        bound = upper-lower
    mode = args.mode

    # Disable dspy caching
    dspy.configure_cache(
    enable_disk_cache=False,
    enable_memory_cache=False,
    )

    prompters = []
    # List of prompters with higher temperature for Self Constistency
    prompters_selfcon = []
    # Evaluate models:
    if "gpt" in args.models:
        if 'selfcon' in args.techs:
            gpt4 = OpenAIPrompter(0.5)
            prompters_selfcon.append(gpt4)
        gpt4 = OpenAIPrompter()
        prompters.append(gpt4)
        print(f'Evaluating on the following GPT-4o model: {gpt4.model_name}')
    if "llama" in args.models:
        if 'selfcon' in args.techs:
            llama = LlamaPrompter(args.new_tok, 0.5, True)
            prompters_selfcon.append(llama)
        llama = LlamaPrompter(args.new_tok)
        prompters.append(llama)
        print(f'Evaluating on the following Llama 3.3 model: {llama.model_name}')
    if "gemma" in args.models:
        if 'selfcon' in args.techs:
            gemma = GemmaPrompter(args.new_tok, 0.5, True)
            prompters_selfcon.append(gemma)
        gemma = GemmaPrompter(args.new_tok)
        prompters.append(gemma)
        print(f'Evaluating on the following Gemma 2 model: {gemma.model_name}')
    # Dspy optimization
    if "dspy" in args.models:
        script_dir = os.path.dirname(os.path.abspath(__file__))
        key = json.load(open(os.path.join(script_dir, 'credentials.json')))['api_key']
        lm = dspy.LM('openai/gpt-4o-mini', api_key=key)
        if 'ts' in args.tasks:
            table_setting_prompter.prompt_dspy(lm, mode)
        if 'to_us' in args.tasks:
            tool_use_prompter.prompt_dspy(lm, mode)
        exit()

    # Decide which tasks to evaluate with which prompting techniques:
    if "ts" in args.tasks:
        if 'base' in args.techs:
            table_setting_prompter.prompt_all_models(prompters, lower, bound)
            print('Finished the Table Setting task (Baseline)')
        if 'rar' in args.techs:
            table_setting_prompter.prompt_all_models_rar(prompters, lower, bound)
            print('Finished the Table Setting task (Rephrase and Respond)')
        if 'meta' in args.techs:
            table_setting_prompter.prompt_all_models_meta(prompters, lower, bound)
            print('Finished the Table Setting task (Metacognitive)')
        if 'selfref' in args.techs:
            table_setting_prompter.prompt_all_models_selfref(prompters, it_selfref, lower, bound)
            print('Finished the Table Setting task (Self Refine)')
        if 'stepback' in args.techs:
            table_setting_prompter.prompt_all_models_stepback(prompters, lower, bound)
            print('Finished the Table Setting task (Stepback)')
        if 'sgicl' in args.techs:
            table_setting_prompter.prompt_all_models_sgicl(prompters, n_ex_sgicl, lower, bound)
            print('Finished the Table Setting task (SG-ICL)')
        if 'contr' in args.techs:
            table_setting_prompter.prompt_all_models_contr(prompters, n_ex_contr, it_selfcon, lower, bound)
            print('Finished the Table Setting task (Contrastive CoT)')
        if 'selfcon' in args.techs:
            table_setting_prompter.prompt_all_models_selfcon(prompters_selfcon, it_selfcon, lower, bound)
            print('Finished the Table Setting task (Self Consistency)')
    if "to_us" in args.tasks:
        if 'base' in args.techs:
            tool_use_prompter.prompt_all_models(prompters, lower, bound)
            print('Finished the Tool Usage task (Baseline)')
        if 'rar' in args.techs:
            tool_use_prompter.prompt_all_models_rar(prompters, lower, bound)
            print('Finished the Tool Usage task (Rephrase and Respond)')
        if 'meta' in args.techs:
            tool_use_prompter.prompt_all_models_meta(prompters, lower, bound)
            print('Finished the Tool Usage task (Metacognitive)')
        if 'selfref' in args.techs:
            tool_use_prompter.prompt_all_models_selfref(prompters, it_selfref, lower, bound)
            print('Finished the Tool Usage task (Self Refine)')
        if 'stepback' in args.techs:
            tool_use_prompter.prompt_all_models_stepback(prompters, lower, bound)
            print('Finished the Tool Usage task (Stepback)')
        if 'sgicl' in args.techs:
            tool_use_prompter.prompt_all_models_sgicl(prompters, n_ex_sgicl, lower, bound)
            print('Finished the Tool Usage task (SG-ICL)')
        if 'contr' in args.techs:
            tool_use_prompter.prompt_all_models_contr(prompters, n_ex_contr, it_selfcon, lower, bound)
            print('Finished the Tool Usage task (Contrastive CoT)')
        if 'selfcon' in args.techs:
            tool_use_prompter.prompt_all_models_selfcon(prompters_selfcon, it_selfcon, lower, bound)
            print('Finished the Tool Usage task (Metacognitive)')
    if "ti_up_mc" in args.tasks:
        if 'base' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models(prompters, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (Baseline)')
        if 'rar' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models_rar(prompters, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (Rephrase and Respond)')
        if 'meta' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models_meta(prompters, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (Metacognitive)')
        if 'selfref' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models_selfref(prompters, it_selfref, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (Self Refine)')
        if 'stepback' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models_stepback(prompters, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (Stepback)')
        if 'sgicl' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models_sgicl(prompters, n_ex_sgicl, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (SG-ICL)')
        if 'contr' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models_contr(prompters, n_ex_contr, it_selfcon, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (Contrastive CoT)')
        if 'selfcon' in args.techs:
            tidy_up_prompter_multi_choice.prompt_all_models_selfcon(prompters_selfcon, it_selfcon, lower, bound)
            print('Finished the Tidy Up task (Multi Choice Questions) (Self Consistency)')
    if "proc_mc" in args.tasks:
        if 'base' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi(prompters, num_runs)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (Baseline)")
        if 'rar' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi_rar(prompters, num_runs)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (Rephrase and Respond)")
        if 'meta' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi_meta(prompters, num_runs)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (Metacognitive)")
        if 'selfref' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi_selfref(prompters, num_runs, it_selfref)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (Self Refine)")
        if 'stepback' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi_stepback(prompters, num_runs)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (Stepback)")
        if 'sgicl' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi_sgicl(prompters, num_runs, n_ex_sgicl)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (SG-ICL)")
        if 'contr' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi_contr(prompters, num_runs, n_ex_contr, it_selfcon)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (Contrastive CoT)")
        if 'selfcon' in args.techs:
            cooking_procedures_prompter.prompt_all_models_multi_selfcon(prompters_selfcon, num_runs, it_selfcon)
            print("Finished the Cooking Procedures Knowledge task (Multi Choice Questions) (Self Consistency)")
    if "mr_mc" in args.tasks:
        if 'base' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models(prompters, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (Baseline)')
        if 'rar' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models_rar(prompters, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (Rephrase and Respond)')
        if 'meta' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models_meta(prompters, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (Metacognitive)')
        if 'selfref' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models_selfref(prompters, it_selfref, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (Self Refine)')
        if 'stepback' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models_stepback(prompters, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (Stepback)')
        if 'sgicl' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models_sgicl(prompters, n_ex_sgicl, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (SG-ICL)')
        if 'contr' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models_contr(prompters, n_ex_contr, it_selfcon, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (Contrastive CoT)')
        if 'selfcon' in args.techs:
            meta_reason_prompter_multi_choice.prompt_all_models_selfcon(prompters_selfcon, it_selfcon, lower, bound)
            print('Finished the Meta-Reasoning task (Multi Choice Questions) (Self Consistency)')

if __name__ == "__main__":
    main()