-
Notifications
You must be signed in to change notification settings - Fork 71
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
7 changed files
with
347 additions
and
70 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,208 @@ | ||
| """ | ||
| This script support LLM API inference with cot/pal/tora prompt. | ||
| It can be used to generate tora corpus. | ||
| Code based on: https://github.com/microsoft/ProphetNet/tree/master/CRITIC | ||
| """ | ||
| import json | ||
| import random | ||
| import os | ||
| import pprint | ||
| import re | ||
| import argparse | ||
| import time | ||
| from datetime import datetime | ||
| from tqdm import tqdm | ||
| from sympy.printing.pretty import pretty | ||
|
|
||
| from api.llm_api import llm_api # use your own API like OpenAI API | ||
| from utils.python_executor import PythonExecutor | ||
| from utils.utils import * | ||
| from utils.parser import * | ||
| # from utils.trajectory import * | ||
| from eval.grader import * | ||
| from utils.data_loader import load_data | ||
| from infer.inference import prepare_data | ||
|
|
||
|
|
||
| def parse_args(): | ||
| parser = argparse.ArgumentParser() | ||
| parser.add_argument("--data_name", default="gsm8k", type=str) | ||
| parser.add_argument("--data_dir", default="./data", type=str) | ||
| parser.add_argument("--model_name_or_path", default="gpt-4", type=str) | ||
| parser.add_argument("--output_dir", default="./output", type=str) | ||
| parser.add_argument("--prompt_type", default="tora", 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("--seed", default=0, type=int) | ||
| parser.add_argument("--start", default=0, type=int) | ||
| parser.add_argument("--end", default=-1, type=int) | ||
| parser.add_argument("--temperature", default=0, type=float) | ||
| parser.add_argument("--n_sampling", default=1, type=int) | ||
| parser.add_argument("--top_p", default=1, type=float) | ||
| parser.add_argument("--max_tokens_per_call", default=1024, type=int) | ||
| parser.add_argument("--shuffle", action="store_true") | ||
| parser.add_argument("--use_train_prompt_format", action="store_true") | ||
| args = parser.parse_args() | ||
| args.top_p = 1 if args.temperature == 0 else args.top_p # top_p must be 1 when using greedy sampling (vllm) | ||
| return args | ||
|
|
||
|
|
||
| def api_with_func_call(engine, prompt, max_tokens, temperature, n, top_p, executor, max_func_call=4, verbose=False): | ||
| if n > 1: | ||
| assert temperature > 0 | ||
|
|
||
| if verbose: | ||
| print("\n======= API with function call (START) =======") | ||
|
|
||
| next_batch_queries = [""] * n | ||
| end_queries = [] | ||
| for i in range(max_func_call): | ||
| batch_outputs = [] | ||
| batch_queries = next_batch_queries | ||
| if len(batch_queries) == 0: | ||
| break | ||
| # get all outputs | ||
| # support batch inference when n > 1 | ||
| if i == 0: | ||
| results = llm_api( | ||
| engine=engine, prompt=prompt + batch_queries[0], max_tokens=max_tokens, temperature=temperature, | ||
| n=n, top_p=top_p, stop=["```output\n", "---"], | ||
| ) | ||
| batch_outputs.extend(results) | ||
| else: | ||
| for k, query in enumerate(batch_queries): | ||
| print("Call {} / {}".format(k+1, len(batch_queries))) | ||
| results = llm_api( | ||
| engine=engine, prompt=prompt + query, max_tokens=max_tokens, temperature=temperature, | ||
| n=1, top_p=top_p, stop=["```output\n", "---"], | ||
| ) | ||
| batch_outputs.append(results[0]) | ||
|
|
||
| # process all outputs | ||
| next_batch_queries = [] | ||
| for query, output in zip(batch_queries, batch_outputs): | ||
| output = output.rstrip() | ||
| query += output | ||
| if verbose: | ||
| print("\n", "-" * 20) | ||
| print(output, end="") | ||
| if "boxed" not in output and output.endswith("```"): | ||
| program = extract_program(query) | ||
| prediction, report = executor.apply(program) | ||
| exec_result = prediction if prediction else report | ||
| exec_result = f"\n```output\n{exec_result.strip()}\n```\n" | ||
| query += exec_result | ||
| if verbose: | ||
| print(exec_result, end="") | ||
| # not end | ||
| if i == max_func_call - 1: | ||
| query += "\nReach max function call limit." | ||
| next_batch_queries.append(query) | ||
| else: | ||
| end_queries.append(query) | ||
|
|
||
| end_queries.extend(next_batch_queries) | ||
| return end_queries | ||
|
|
||
|
|
||
|
|
||
| def main(args): | ||
| examples, processed_samples, out_file = prepare_data(args) | ||
| # init python executor | ||
| if "pal" in args.prompt_type: | ||
| executor = PythonExecutor(get_answer_expr='solution()') | ||
| else: | ||
| executor = PythonExecutor(get_answer_from_stdout=True) | ||
|
|
||
| writer = open(out_file, 'w') | ||
| correct, wrong = 0, 0 | ||
|
|
||
| for example in tqdm(examples, total=len(examples)): | ||
| idx = example['idx'] | ||
|
|
||
| # parse question and answer | ||
| example['question'] = parse_question(example, args.data_name) | ||
| gt_cot, gt_ans = parse_ground_truth(example, args.data_name) | ||
| full_prompt = construct_prompt(args, example) | ||
|
|
||
| # call LLM, return list | ||
| if "tora" in args.prompt_type: | ||
| results = api_with_func_call( | ||
| engine=args.model_name_or_path, | ||
| prompt=full_prompt, | ||
| max_tokens=args.max_tokens_per_call, | ||
| temperature=args.temperature, | ||
| n=args.n_sampling, | ||
| top_p=args.top_p, | ||
| executor=executor, | ||
| ) | ||
| else: | ||
| stop_tokens = ["</s>", "---", "```output"] | ||
| if args.prompt_type in ['cot']: | ||
| stop_tokens.append("\n\n") | ||
| results = llm_api( | ||
| engine=args.model_name_or_path, | ||
| prompt=full_prompt, | ||
| max_tokens=args.max_tokens_per_call, | ||
| temperature=args.temperature, | ||
| n=args.n_sampling, | ||
| top_p=args.top_p, | ||
| stop=stop_tokens, | ||
| ) | ||
| # deal with error | ||
| if results == ['error']: | ||
| print(">>> Error API call") | ||
| continue | ||
| print("Get {} results".format(len(results))) | ||
| # get prediction | ||
| predictions = [] | ||
| reports = [] | ||
| for r in results: | ||
| pred, report = run_execute(executor, r, args.prompt_type, execute=True) | ||
| predictions.append(pred) | ||
| reports.append(report) | ||
| print("Executed {} results".format(len(predictions))) | ||
|
|
||
| scores = [math_equal(p, gt_ans, timeout=True) for p in predictions] | ||
|
|
||
| is_correct = scores[0] | ||
| if is_correct: | ||
| correct += 1 | ||
| else: | ||
| wrong += 1 | ||
|
|
||
| sample = {'idx': idx, 'question': example['question'], 'gt_cot': gt_cot, 'gt': gt_ans, | ||
| 'pred': predictions, 'score': scores} | ||
|
|
||
| if args.prompt_type == "cot": | ||
| sample.update({'code': results}) | ||
| elif "tora" in args.prompt_type or "pal" in args.prompt_type: | ||
| sample.update({'report': reports, 'code': results}) | ||
| # add remain fields | ||
| for key in ['level', 'type', 'unit', 'solution_type', 'choices', 'solution', 'ques_type', \ | ||
| 'ans_type', 'answer_type', 'dataset', 'subfield', 'filed', 'theorem', 'answer']: | ||
| if key in example: | ||
| sample[key] = example[key] | ||
|
|
||
| print(idx) | ||
| show_sample(sample) | ||
| if correct + wrong > 0: | ||
| print("Avg Acc:", correct / (correct + wrong)) | ||
| print() | ||
|
|
||
| try: | ||
| writer.write(json.dumps(sample) + '\n') | ||
| writer.flush() | ||
| except: | ||
| print(">>> Error writing to file") | ||
| continue | ||
|
|
||
| writer.close() | ||
| print() | ||
| print(correct / (correct + wrong)) | ||
|
|
||
|
|
||
| if __name__ == "__main__": | ||
| args = parse_args() | ||
| set_seed(args.seed) | ||
| main(args) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -1,29 +1,29 @@ | ||
| set -ex | ||
|
|
||
| # MODEL_NAME_OR_PATH="llm-agents/tora-code-34b-v1.0" | ||
| MODEL_NAME_OR_PATH="llm-agents/tora-70b-v1.0" | ||
| MODEL_NAME_OR_PATH="llm-agents/tora-code-34b-v1.0" | ||
| # MODEL_NAME_OR_PATH="llm-agents/tora-70b-v1.0" | ||
|
|
||
| # DATA_LIST = ['math', 'gsm8k', 'gsm-hard', 'svamp', 'tabmwp', 'asdiv', 'mawps'] | ||
|
|
||
| DATA="math" | ||
| # DATA="gsm8k" | ||
| DATA_NAME="math" | ||
| # DATA_NAME="gsm8k" | ||
|
|
||
| SPLIT="test" | ||
| PROMPT_TYPE="tora" | ||
| NUM_TEST_SAMPLE=-1 | ||
|
|
||
|
|
||
| CUDA_VISIBLE_DEVICES=2,3 TOKENIZERS_PARALLELISM=false \ | ||
| python -m infer.inference \ | ||
| CUDA_VISIBLE_DEVICES=0 TOKENIZERS_PARALLELISM=false \ | ||
| python -um infer.inference \ | ||
| --model_name_or_path ${MODEL_NAME_OR_PATH} \ | ||
| --data ${DATA} \ | ||
| --data_name ${DATA_NAME} \ | ||
| --split ${SPLIT} \ | ||
| --prompt_type ${PROMPT_TYPE} \ | ||
| --use_train_prompt_format \ | ||
| --num_test_sample ${NUM_TEST_SAMPLE} \ | ||
| --seed 0 \ | ||
| --temperature 0 \ | ||
| --n_sampling 1 \ | ||
| --top_p 0.95 \ | ||
| --top_p 1 \ | ||
| --start 0 \ | ||
| --end -1 \ |
Oops, something went wrong.