File size: 7,809 Bytes
d7970b4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import pandas as pd
import random
import copy
from tqdm import tqdm
import time
import json
from openai import OpenAI
import os
import dotenv
import tempfile
import numpy as np

from typing import List, Dict, Any, Optional
from dataclasses import dataclass
from concurrent.futures import ThreadPoolExecutor
from huggingface_hub import login
from transformers import AutoModelForCausalLM, AutoTokenizer
dotenv.load_dotenv()

API_KEY = ""
login(token=API_KEY)
huggingface_models=["meta-llama/Llama-3.1-8B-Instruct","Qwen/Qwen2.5-7B-Instruct","Qwen/Qwen2.5-14B-Instruct","Qwen/Qwen2.5-7B-Instruct"]
model_name = "Model_name" # Change here for testing your model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

class GPTReranker:
    def __init__(self, api_key: str, model_name: str = model_name, window_size: int = 30, overlap: int = 10):
        if window_size <= overlap:
            raise ValueError("Window size must be greater than overlap")
        if overlap < 0:
            raise ValueError("Overlap must be non-negative")
            
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForCausalLM.from_pretrained(
            model_name,
            torch_dtype=torch.float16,
            device_map="auto"
        )
        self.model.gradient_checkpointing_enable()
        self.model.eval()  # Set to evaluation mode
        
        self.window_size = window_size
        self.overlap = overlap
        
    def _create_messages(self, query: str, passages: List[str], start_idx: int) -> str:
        prompt = f"""<s>[INST] You are an expert that ranks passages based on their relevance to a given query. 
The most relevant passage should be ranked first. 
Important: Do not just sort the passage numbers. Evaluate each passage's content for relevance.

Query: {query}

I will give you passages numbered from {start_idx+1} to {start_idx+len(passages)}. Rank them by relevance to the query, with the most relevant first.

"""
        for i, passage in enumerate(passages):
            prompt += f"[{start_idx+i+1}] {passage}\n"
        
        prompt += """
Based on the content of each passage (not just their numbers), rank them from most to least relevant. 
Format: [most_relevant] > [next] > [next]. No explanation needed.[/INST]"""
        
        return prompt

    def get_ranking_for_group(self, query: str, passages: List[str], start_idx: int = 0, max_retries: int = 3) -> List[int]:
        prompt = self._create_messages(query, passages, start_idx)
        
        for attempt in range(max_retries):
            try:
                inputs = self.tokenizer(prompt, return_tensors="pt").to(self.model.device)
                
                with torch.no_grad():
                    outputs = self.model.generate(
                        **inputs,
                        max_new_tokens=150,
                        temperature=0.0,
                        do_sample=False,
                        pad_token_id=self.tokenizer.eos_token_id
                    )
                ranking_str = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
                ranking_str = ranking_str[len(prompt):]
                
                raw_ranks = Evaluator.clean_ranking_response(ranking_str)
                global_ranks = []
                for rank in raw_ranks:
                    local_idx = rank - (start_idx + 1)
                    if 0 <= local_idx < len(passages):
                        global_idx = start_idx + local_idx
                        global_ranks.append(global_idx)
                
                return global_ranks
                
            except Exception as e:
                print(f"Attempt {attempt + 1} failed: {str(e)}")
                if attempt == max_retries - 1:
                    raise
                time.sleep(5)

    def get_ranking(self, query: str, passages: List[str]) -> str:
        if not passages:
            raise ValueError("No passages provided")
        
        first_group = passages[:40]
        second_group = passages[40:80]
        third_group = passages[80:]
        
        first_ranks = self.get_ranking_for_group(query, first_group, 0)
        print(f"First group top 10: {first_ranks[:10]}")
        second_ranks = self.get_ranking_for_group(query, second_group, 40)
        print(f"Second group top 10: {second_ranks[:10]}")
        third_ranks = self.get_ranking_for_group(query, third_group, 80)
        print(f"Third group top 10: {third_ranks[:10]}")
        
        top_30_indices = []
        if first_ranks:
            top_30_indices.extend(first_ranks[:10])
        if second_ranks:
            top_30_indices.extend(second_ranks[:10])
        if third_ranks:
            top_30_indices.extend(third_ranks[:10])
        top_30_passages = [passages[i] for i in top_30_indices]
        
        final_local_ranks = self.get_ranking_for_group(query, top_30_passages, 0)
        
        final_indices = []
        for rank in final_local_ranks:
            if rank < len(top_30_indices):
                final_indices.append(top_30_indices[rank])
        
        remaining_top = [idx for idx in top_30_indices if idx not in final_indices]
        final_indices.extend(remaining_top)
        
        all_other_indices = [i for i in range(len(passages)) if i not in top_30_indices]
        final_indices.extend(all_other_indices)
        
        ranking_str = " > ".join(f"[{r+1}]" for r in final_indices)
        return ranking_str
@dataclass
class RankingResult:
    query: str
    correct_passage: str
    ranking: str
    correct_idx: int
    passages: List[str]
    ranks: List[int]

class Evaluator:
    @staticmethod
    def clean_ranking_response(response: str) -> List[int]:
        return [int(num) for num in ''.join(c if c.isdigit() else ' ' for c in response).split()]

def process_query(row: pd.Series, reranker: GPTReranker) -> Optional[RankingResult]:
    try:
        query = row['query']
        correct_passage_idx = int(row['correct_passage_index'])
        passages = [row[f'passage_{i}'] for i in range(1, 101)]
        
        ranking_response = reranker.get_ranking(query, passages)
        ranks = [i-1 for i in Evaluator.clean_ranking_response(ranking_response)]
        
        return RankingResult(
            query=query,
            correct_passage=passages[correct_passage_idx - 1],
            ranking=ranking_response,
            correct_idx=correct_passage_idx,
            passages=passages,
            ranks=ranks
        )
    except Exception as e:
        print(f"Error processing query: {str(e)}")
        return None

def save_results(results: List[RankingResult], filename: str):
    results_data = []
    for result in results:
        results_data.append({
            'query': result.query,
            'correct_passage': result.correct_passage,
            'ranking': result.ranking,
            'correct_idx': result.correct_idx,
            'passages': result.passages,
            'ranks': result.ranks
        })
    
    with open(filename, 'w', encoding='utf-8') as f:
        json.dump(results_data, f, ensure_ascii=False, indent=2)


def main():
    df = pd.read_csv('./datasets/QACandidate_Pool.csv')
    reranker = GPTReranker(API_KEY)
    results = []
    for _, row in tqdm(df.iterrows()):
        if result := process_query(row, reranker):
            print(f"\nQuery: {result.query}") # Show the query sentence
            print(f"Correct index: {result.correct_idx}") # Show the correct index
            print(f"Ranks: {result.ranks[:10]}")  # Show first 10 ranks
            results.append(result)
            time.sleep(1) 

            save_results(results, 'output.json') # Change here for your output file name
            time.sleep(1)
    
main()