File size: 20,736 Bytes
1a0cf07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
import argparse
import datetime
import json
import logging
import multiprocessing
import os
import re
from abc import ABC, abstractmethod

import hjson
import numpy as np
import openai
from tqdm import tqdm
from sklearn.metrics.pairwise import cosine_similarity

from data_loader import load_data
from code_executor import PythonExecutor
from utils import (Agent, LLMClient, PromptTemplate, api_configs,
                   extract_and_parse_markup, setup_logging)
from data_utils import parse_question, parse_ground_truth
from evaluate import evaluate


logger = setup_logging()

class RetrievalAugmentation:
    # TODO: implement the retrieval augmentation later 
    def __init__(self, dataset, embeddings):
        self.dataset = dataset
        self.embeddings = embeddings

    def get_similar_examples(self, query_embedding, n=3):
        similarities = cosine_similarity([query_embedding], self.embeddings)[0]
        top_indices = similarities.argsort()[-n:][::-1]
        return [self.dataset[i] for i in top_indices]

class SwiftAgent(Agent):
    def __init__(self, prompt_template, llm_client, retrieval_augmentation=None):
        super().__init__(prompt_template, llm_client)
        self.retrieval_augmentation = retrieval_augmentation
        self.plans = {}
        self.codes = {}

    def generate_response(self, prompt, reasoning, current_solution, plan, critical_feedback, prefill=True):
        logger.info("SwiftAgent generating response")
        if self.retrieval_augmentation:
            query_embedding = self.get_query_embedding(prompt)
            similar_examples = self.retrieval_augmentation.get_similar_examples(query_embedding)
            examples_text = "\n".join(similar_examples) # TODO: add more context to the prompt
        else:
            examples_text = "No similar examples available."
        
        swift_prompt = self.prompt_template.format(
            "swift",
            prompt=prompt,
            current_reasoning=reasoning, # TODO: check if this is needed
            examples=examples_text,
            current_solution=current_solution,
            critical_feedback=critical_feedback,
            revised_plan=plan
        )
        # logger.info(f"SwiftAgent prompt:\n{swift_prompt}")

        messages = [
            {"role": "system", "content": ''},
            {"role": "user", "content": swift_prompt}
        ]
        if prefill:
            messages.append({"role": "assistant", "content": "<plan>"}) # prefix-filling 
        
        response = self.llm_client.generate_response(messages) 
        if prefill:
            response = "<plan>" + response
        
        try:
            parsed_response = extract_and_parse_markup(response)
            return parsed_response
        except json.JSONDecodeError:
            logger.error("Error: Swift's response was not in valid JSON format. Returning raw response.")
            return response

    def get_query_embedding(self, query):
        # Implement query embedding generation
        return np.random.rand(768)  # Placeholder, replace with actual embedding
 
class SageAgent(Agent):
    def __init__(self, prompt_template, llm_client):
        super().__init__(prompt_template, llm_client)
        self.feedbacks = {}
        self.plans = {}
        

    def generate_response(self, prompt, reasoning, current_solution, prefill=True):
        logger.info("SageAgent generating response")
        sage_prompt = self.prompt_template.format(
            "sage",
            prompt=prompt,
            reasoning=reasoning, 
            current_solution=current_solution
        )
        # logger.info(f"SageAgent prompt:\n{sage_prompt}")
        
        messages = [
            {"role": "system", "content": ""},
            {"role": "user", "content": sage_prompt}
        ]
        if prefill:
            messages.append({"role": "assistant", "content": "<solved>"}) # prefix-filling 
        
        response = self.llm_client.generate_response(messages)
        # logger.info(f"SageAgent raw response:\n{response}")
        if prefill:
            response = "<solved>" + response
        try:
            parsed_response = extract_and_parse_markup(response)
            return parsed_response
        except json.JSONDecodeError:
            logger.error("Error: Sage's response was not in valid JSON format. Returning raw response.")
            return response

class RewardModel:
    def __init__(self, prompt_template, llm_client):
        self.prompt_template = prompt_template
        self.llm_client = llm_client 
        self.scores = [] 
        self.feedbacks = [] 
        self.stagnant_count = 0

    def calculate_reward(self, problem, reasoning, current_solution, prefill=True):
        reward_prompt = self.prompt_template.format(
            "reward",
            problem=problem,
            reasoning= reasoning,
            current_solution=current_solution
        )
        # logger.info(f"RewardModel prompt:\n{reward_prompt}")
        
        messages = [
            {"role": "system", "content": ""},
            {"role": "user", "content": reward_prompt}
        ]
        if prefill:
            messages.append({"role": "assistant", "content": "<feedback>"}) # prefix-filling 
        
        reward_response = self.llm_client.generate_response(messages) 
        if prefill:
            reward_response = "<feedback>" + reward_response
        
        try:
            parsed_response = extract_and_parse_markup(reward_response)
            score = int(parsed_response["score"])
            
            # Update stagnant_count based on score comparison
            if len(self.scores) > 0 and score <= self.scores[-1]:
                self.stagnant_count += 1
            else:
                self.stagnant_count = 0
            
            return parsed_response
        except json.JSONDecodeError:
            logger.error("Error: Reward model's response was not in valid JSON format. Returning raw response.")
            return reward_response

    def should_consult_sage(self):
        # This method remains unchanged
        return self.stagnant_count >= 1 or (len(self.scores) > 0 and self.scores[-1] < 5)

class SwiftSage:
    def __init__(self, dataset, embeddings, prompt_template_dir, swift_config, sage_config, reward_config, use_retrieval=True, start_with_sage=False):
        prompt_template = PromptTemplate(prompt_template_dir)
        retrieval_augmentation = RetrievalAugmentation(dataset, embeddings) if use_retrieval else None
        
        # add logger to the following LLMClient 
        swift_llm = LLMClient(**swift_config, logger=logger)
        sage_llm = LLMClient(**sage_config, logger=logger)
        reward_llm = LLMClient(**reward_config, logger=logger)

        self.swift = SwiftAgent(prompt_template, swift_llm, retrieval_augmentation)
        self.sage = SageAgent(prompt_template, sage_llm)
        self.reward_model = RewardModel(prompt_template, reward_llm)
        self.start_with_sage = start_with_sage
        # self.executor = PythonExecutor(get_answer_from_stdout=True)
    
    def solve(self, problem, max_iterations=10, reward_threshold=8):
        logger.info(f"Starting to solve problem: {problem}")
        current_solution = "No current solution yet." # final answer
        current_reasoning = "No reasoning steps yet." # reasoning steps
        plan = "Initial plan: Take a deep breath and think step by step."
        critical_feedback = "No critical feedback yet."  # Initialize critical_feedback
        solved = False
        for i in range(max_iterations):
            logger.info(f"Iteration {i+1}")
            

            #  Use the Sage Agent 
            if (i == 0 and self.start_with_sage) or self.reward_model.should_consult_sage():
                sage_parsed = self.sage.generate_response(problem, current_reasoning, current_solution) 
                critical_feedback = sage_parsed["critical_feedback"]
                # plan = "\n - " + "\n - ".join(sage_parsed["revised_plan"])
                current_reasoning = sage_parsed["reasoning_steps"] 
                current_code = sage_parsed["code"] 

                solved = sage_parsed["solved"].lower() == "true" if i != 0 else sage_parsed["solved"] 
                if solved:
                    return current_reasoning, current_solution
                logger.info(f"Sage's feedback (iteration {i+1}):\n{critical_feedback}")
                # logger.info(f"Sage's reasoning steps:\n{current_reasoning}")
                self.sage.feedbacks[i] = critical_feedback
                
                # run the code 
                executor = PythonExecutor(get_answer_from_stdout=True)
                code_result, code_report = executor.apply(current_code)
                logger.info(f"Sage Code execution report: {code_report}")
                logger.info(f"Sage Code execution result: {code_result}")
                current_reasoning = current_reasoning + f"\n\nThe generated code is:\n\n```python\n{current_code}\n```"
                current_solution = "Answer (from running the code):\n " + code_result
                
                # current_solution = sage_parsed["final_answer"]
                logger.info("Activated Sage, so we should return the reasoning and solution from Sage.")
                return current_reasoning, current_solution
            
            if not solved:
                # Use the Swift Agent 
                swift_parsed = self.swift.generate_response(problem, current_reasoning, current_solution, plan, critical_feedback)
                
                if "code" not in swift_parsed and "final_answer" not in swift_parsed: 
                    logger.info("Swift's response does not contain the 'final_answer' or 'code' field. Returning raw response.")
                    self.reward_model.scores.append(0)
                    self.reward_model.feedbacks.append("No feedback")
                    self.reward_model.stagnant_count += max_iterations # force to use Sage Agent
                    continue 
                
                current_plan = swift_parsed["plan"]
                current_code = swift_parsed["code"]
                current_answer = swift_parsed.get("final_answer", None)

                self.swift.plans[i] = current_plan
                self.swift.codes[i] = current_code  

                logger.info(f"Swift's plan:\n{current_plan}")
                logger.info(f"Swift's code:\n{current_code}")

                # Call sandbox to run the code and get the result
                executor = PythonExecutor(get_answer_from_stdout=True)
                code_result, code_report = executor.apply(current_code)
                logger.info(f"Code execution report: {code_report}")
                logger.info(f"Code execution result: {code_result}")
            
                current_reasoning = current_plan + f"\nThe generated code is:\n```python\n{current_code}\n```"
                current_solution = "Answer (from running the code):\n " + code_result

                # Calling the reward model to provide feedback and score 
                reward_parsed = self.reward_model.calculate_reward(problem, current_reasoning, current_solution)
                score = int(reward_parsed["score"])
                feedback = reward_parsed["feedback"] 
                prev_score = self.reward_model.scores[-1] if len(self.reward_model.scores) > 0 else 0
                self.reward_model.scores.append(score)
                self.reward_model.feedbacks.append(feedback)

                # detect if the score is lower than the previous score
                logger.info(f"Reward for iteration {i+1}: {score}/10")
                logger.info(f"Feedback: {feedback}")

                if False and score < prev_score:
                    logger.info("Score is lower than the previous score. Stopping the iteration. Reverting to the previous solution and reasoning.")
                    # revert to the previous solution and reasoning
                    current_solution = self.swift.codes[i-1]
                    current_reasoning = self.swift.plans[i-1]
                    continue 

                
                critical_feedback = feedback 

            
            if score >= reward_threshold or solved:
                logger.info("Perfect solution found!")
                return current_reasoning, current_solution 
            
            
            if self.reward_model.should_consult_sage():
                logger.info("Reward model: The solution quality hasn't improved recently. Consulting Sage for the next iteration.")
        
        logger.info("Max iterations reached without finding a perfect solution.")
        logger.info("Problem solving completed")
        return current_reasoning, current_solution


def run_test(swiftsage, problem, max_iterations=5, reward_threshold=8):
    logger.info(f"Testing problem: {problem}")
    reasoning, solution = swiftsage.solve(problem, max_iterations, reward_threshold)
    logger.info(f"Final reasoning:\n{reasoning}")
    logger.info(f"Final solution:\n{solution}")
    logger.info("=" * 50)


def run_benchmark(swiftsage, args, max_iterations=5, reward_threshold=8):
    examples = load_data(args.dataset_name, args.split, args.data_dir, args.num_test_sample)

    res = []
    skip_ids = []

    output_path = os.path.join(args.output_path, f"{args.dataset_name}.jsonl")
    if os.path.exists(output_path):
        with open(output_path) as fr:
            model_responses = fr.readlines()

        for item in model_responses:
            item = json.loads(item)
            res.append(item)
            skip_ids.append(item["idx"])

    for example in tqdm(examples, desc=args.dataset_name):
        if example["idx"] in skip_ids:
            continue
        question = parse_question(example, args.dataset_name)
        gt_ans = parse_ground_truth(example, args.dataset_name)
        reasoning, solution = swiftsage.solve(question, max_iterations, reward_threshold)
        
        # TODO: extract answer from solution

        cur_res = {
            "idx": example["idx"],
            "question": question,
            "gt": gt_ans,
            "pred": solution,
            "reasoning": reasoning,
        }
        res.append(cur_res)

        with open(output_path, "a") as fw:
            fw.write(json.dumps(res[-1]) + "\n")
    
    # Evaluate the results
    res, result_metric = evaluate(res)
    with open(args.output_path, f"{args.dataset_name}_score.jsonl", "w") as fw:
        for item in res:
            fw.write(json.dumps(item) + "\n")
    with open(args.output_path, f"{args.dataset_name}_metric.jsonl", "w") as fw:
        fw.write(json.dumps(result_metric) + "\n")


def main(args):

    # TODO: for retrieval augmentation (not implemented yet now) 
    # dataset = ["Example problem 1: ...", "Example problem 2: ...", "Example problem 3: ..."]
    # embeddings = np.random.rand(len(dataset), 768)  # Placeholder, replace with actual embeddings


    # Configuration for each LLM
    # swift_config = {
    #     "model_id": "Meta-Llama-3.1-8B-Instruct",
    #     "api_config": api_configs['SambaNova']
    # }

    # reward_config = {
    #     "model_id": "Meta-Llama-3.1-70B-Instruct",
    #     "api_config": api_configs['SambaNova']
    # }

    # sage_config = {
    #     "model_id": "Meta-Llama-3.1-405B-Instruct",
    #     "api_config": api_configs['SambaNova']
    # }

    swift_config = {
        "model_id": args.swift_model_id,
        "api_config": api_configs[args.api_provider]
    }

    reward_config = {
        "model_id": args.reward_model_id,
        "api_config": api_configs[args.api_provider]
    }

    sage_config = {
        "model_id": args.sage_model_id,
        "api_config": api_configs[args.api_provider]
    }

    # specify the path to the prompt templates
    prompt_template_dir = args.prompt_template_dir
    dataset = [] 
    embeddings = [] # TODO: for retrieval augmentation (not implemented yet now)
    s2 = SwiftSage(
        dataset,
        embeddings,
        prompt_template_dir,
        swift_config,
        sage_config,
        reward_config,
        use_retrieval=args.use_retrieval,
        start_with_sage=args.start_with_sage,
    )

    if args.eval_mode == "test":
        test_problems = [
            "Solve the equation: 2x + 5 = 13", # 0
            "If h(x)=x-4 and g(h(x))=x^2-8x+10, find g(x)? show the formula for g(x)", # 1
            "Solve the equation: 6y + 5 = 29", # 2
            "Who lives longer, Lowell Sherman or Jonathan Kaplan?", # 3
            "9.9 or 9.11 --  which is bigger?", # 4
            "How can you solve the quadratic equation 3x^2 + 7x + 4 = 0 using the quadratic formula?", # 5
            "Explain why sound waves cannot travel in a vacuum?", # 6
            "How many grams of hydrogen (H) are present in 23.5 grams of water (H2O)?", # 7
            "What is the distance between the points (2, 3) and (5, 8)?", # 8
            "Why can the Hubble telescope capture clear images of distant stars and galaxies, but not a detailed image of Pluto?", # 9
            """A rectangular band formation is a formation with $m$ band members in each of $r$ rows, where $m$ and $r$ are integers. A particular band has less than 100 band members. The director arranges them in a rectangular formation and finds that he has two members left over. If he increases the number of members in each row by 1 and reduces the number of rows by 2, there are exactly enough places in the new formation for each band member. What is the largest number of members the band could have?""",
            """Tim wants to invest some money in a bank which compounds quarterly with an annual interest rate of $7\%$. To the nearest dollar, how much money should he invest if he wants a total of $\$60,\!000$ at the end of $5$ years?""",
            """In an SR latch built from NOR gates, which condition is not allowed

            Options:
            [ "S=0, R=2", "S=2, R=2", "S=1, R=1", "S=1, R=-1", "S=1, R=2", "S=0, R=0", "S=2, R=0", "S=1, R=0", "S=2, R=1", "S=0, R=1" ]

            Which one is the correct answer?""",
            # ... add other problems here ...
            """How many letter r are there in the word "strawberry"?"""
        ]

        # for problem in test_problems:
        pid = 7
        print(f"Problem {pid}: {test_problems[pid]}")
        run_test(s2, test_problems[pid], args.max_iterations, args.reward_threshold)
    elif args.eval_mode == "benchmark":
        run_benchmark(s2, args, args.max_iterations, args.reward_threshold)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("--eval_mode", default="test", choices=["test", "benchmark"], type=str)

    parser.add_argument("--dataset_name", default="MATH", type=str)
    parser.add_argument("--data_dir", default="./data", type=str)
    parser.add_argument("--split", default="test", type=str)
    parser.add_argument("--num_test_sample", default=-1, type=int)  # -1 for full data

    parser.add_argument("--api_provider", default="Together", choices=["Together", "SambaNova"], type=str)
    parser.add_argument("--swift_model_id", default="meta-llama/Meta-Llama-3-8B-Instruct-Turbo", type=str)
    parser.add_argument("--reward_model_id", default="meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo", type=str)
    parser.add_argument("--sage_model_id", default="meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo", type=str)

    parser.add_argument("--prompt_template_dir", default='./prompt_templates', type=str)
    parser.add_argument("--use_retrieval", action="store_true")
    parser.add_argument("--start_with_sage", action="store_true")

    parser.add_argument("--max_iterations", default=5, type=int)
    parser.add_argument("--reward_threshold", default=8, type=int)

    parser.add_argument("--save_outputs", action="store_true")
    parser.add_argument("--output_path", default="./output", type=str)
    parser.add_argument("--overwrite", action="store_true")

    args = parser.parse_args()

    # remove console output for benchmark evaluation
    if args.eval_mode != "test":
        root_logger = logging.getLogger("")
        for handler in root_logger.handlers:
            if isinstance(handler, logging.StreamHandler):
                root_logger.removeHandler(handler)
                break

    if args.api_provider == "SambaNova":
        args.swift_model_id = args.swift_model_id.split("/")[-1][:-len("Turbo")]
        args.reward_model_id = args.reward_model_id.split("/")[-1][:-len("Turbo")]
        args.sage_model_id = args.sage_model_id.split("/")[-1][:-len("Turbo")]

    multiprocessing.set_start_method('spawn')
    main(args)