MTEB Leaderboard Chinese Reranking Results
We have validated the performance of our model on the mteb-chinese-reranking leaderboard. Currently, the open-source models on this leaderboard are primarily bidirectional discriminative models (BERT-like models). The only unidirectional generative model (GPT-like model) is gte-Qwen1.5-7B-instruct, which has an average score of 66.38, ranking 25th, with less than ideal results. Our self-developed unidirectional generative model, 360Zhinao-1.8B-reranking, achieved an average score of 70.13, currently ranking first overall and first among open-source models, opening up new possibilities for generative models to undertake discriminative tasks. For more detail, please check 360zhinao.
Model | T2Reranking | MMarcoReranking | CMedQAv1 | CMedQAv2 | Avg |
---|---|---|---|---|---|
360Zhinao-1.8B-Reranking | 68.55 | 37.29 | 86.75 | 87.92 | 70.13 |
piccolo-large-zh-v2 | 67.15 | 33.39 | 90.14 | 89.31 | 70 |
Baichuan-text-embedding | 67.85 | 34.3 | 88.46 | 88.06 | 69.67 |
stella-mrl-large-zh-v3.5-1792d | 66.43 | 28.85 | 89.18 | 89.33 | 68.45 |
PEG | 69.43 | 33.55 | 86.56 | 84.09 | 68.41 |
bge-reranker-base | 67.28 | 35.46 | 81.27 | 84.1 | 67.03 |
bge-reranker-large | 67.6 | 37.17 | 82.14 | 84.19 | 67.78 |
Requirements
pip install -r requirements.txt
If your GPU supports fp16 or bf16 precision, we also recommend installing flash-attention (now with support for flash attention 2) to improve your runtime efficiency and reduce memory usage. (flash-attention is optional and not required for running this project)
git clone https://github.com/Dao-AILab/flash-attention
cd flash-attention && pip install .
# The installation below is optional and might be slow.
# pip install csrc/layer_norm
# No need to install the following if the flash-attn version is above 2.1.1.
# pip install csrc/rotary
You can also use the following command to install flash-attention.
FLASH_ATTENTION_FORCE_BUILD=TRUE ./miniconda3/bin/python -m pip install flash-attn==2.3.6
Model Introduction
The 360Zhinao-1.8B-Reranking model utilizes the self-developed 360Zhinao-1.8b-Base model as its foundation. Through iterative discovery and resolution of the following technical issues, it continuously stimulates the world knowledge inherent in the large model during the pre-training phase, better bridging the gap between generative models and discriminative tasks.
Data Processing
The model training did not utilize world knowledge, meaning it neither continue pre-training with domain-specific data nor fine-tuned datasets outside of the four datasets on the leaderboard. It only used the four datasets within the leaderboard, carefully iterating through data perception, data cleaning and data mining to ensure that the ranking in individual tasks could reach the top three level.
Resolving Task Conflicts and Catastrophic Forgetting
When merging four tasks, due to different data domain distributions, answer patterns, training data volumes, convergence steps, and even sequence lengths, conflicts exist between different tasks. Deeply resolving these conflict issues is crucial to obtaining a universal model with the best comprehensive indicators across different tasks.
Resolving Training Instability
Unlike generative tasks that produce multiple characters, using generative models for discriminative tasks requires the model to output a continuous value. Therefore, there is an oscillation problem during the training process. Deeply analyzing and resolving training instability can result in a model with better generalization and robustness.
Inference Script
You can copy the following scripts to mteb-eval-scripts, then replace FlagReranker with FlagRerankerCustom in eval_cross_encoder scripts, then run eval_cross_encoder to reproduce our complete performance on the mteb-chinese-reranking leaderboard.
from typing import cast, List, Union, Tuple, Dict, Optional
import numpy as np
import torch
from tqdm import tqdm
from transformers import AutoModel, AutoTokenizer, AutoModelForSequenceClassification
import transformers
from transformers.trainer_pt_utils import LabelSmoother
IGNORE_TOKEN_ID = LabelSmoother.ignore_index
def preprocess(
sources,
tokenizer: transformers.PreTrainedTokenizer,
max_len: int = 1024,
system_message: str = "",
device = None,
) -> Dict:
roles = {"user": "<|im_start|>user", "assistant": "<|im_start|>assistant"}
answer_len = 64
im_start = tokenizer.im_start_id
im_end = tokenizer.im_end_id
nl_tokens = tokenizer('\n').input_ids
_system = tokenizer('system').input_ids + nl_tokens
_user = tokenizer('user').input_ids + nl_tokens
_assistant = tokenizer('assistant').input_ids + nl_tokens
# Apply prompt templates
input_ids, targets = [], []
for i, source in enumerate(sources):
## system_message
input_id, target = [], []
system = [im_start] + _system + tokenizer(system_message, max_length=max_len-answer_len, truncation=True).input_ids + [im_end] + nl_tokens
input_id += system
target += [im_start] + [IGNORE_TOKEN_ID] * (len(system)-3) + [im_end] + nl_tokens
assert len(input_id) == len(target)
## query ans
source = "\n\n".join(source)
role = "<|im_start|>user"
_input_id = tokenizer(role, max_length=max_len-answer_len, truncation=True).input_ids + nl_tokens + \
tokenizer(source, max_length=max_len-answer_len, truncation=True).input_ids + [im_end] + nl_tokens
input_id += _input_id
if role == '<|im_start|>user':
_target = [im_start] + [IGNORE_TOKEN_ID] * (len(_input_id)-3) + [im_end] + nl_tokens
elif role == '<|im_start|>assistant':
_target = [im_start] + [IGNORE_TOKEN_ID] * len(tokenizer(role, max_length=max_len-answer_len, truncation=True).input_ids) + \
_input_id[len(tokenizer(role, max_length=max_len-answer_len, truncation=True).input_ids)+1:-2] + [im_end] + nl_tokens
else:
raise NotImplementedError
target += _target
## label use placeholder 0; It will be masked later in the modeling_zhinao.py
role = "<|im_start|>assistant"
_input_id = tokenizer(role, max_length=max_len-answer_len, truncation=True).input_ids + nl_tokens + \
tokenizer("0", max_length=max_len-answer_len, truncation=True).input_ids + [im_end] + nl_tokens
input_id += _input_id
if role == '<|im_start|>user':
_target = [im_start] + [IGNORE_TOKEN_ID] * (len(_input_id)-3) + [im_end] + nl_tokens
elif role == '<|im_start|>assistant':
_target = [im_start] + [IGNORE_TOKEN_ID] * len(tokenizer(role, max_length=max_len-answer_len, truncation=True).input_ids) + \
_input_id[len(tokenizer(role, max_length=max_len-answer_len, truncation=True).input_ids)+1:-2] + [im_end] + nl_tokens
else:
raise NotImplementedError
target += _target
assert len(input_id) == len(target)
input_id += [tokenizer.pad_token_id] * (max_len - len(input_id))
target += [IGNORE_TOKEN_ID] * (max_len - len(target))
if len(input_id) > max_len:
print("max_len_error")
print(tokenizer.decode(input_id))
input_ids.append(input_id[:max_len])
targets.append(target[:max_len])
input_ids = torch.tensor(input_ids, dtype=torch.int)
targets = torch.tensor(targets, dtype=torch.int)
#print(f"input_ids {input_ids.shape}")
#print(f"targets {targets.shape}")
return dict(
input_ids=input_ids.to(device),
labels=targets.to(device),
attention_mask=input_ids.ne(tokenizer.pad_token_id).to(device),
)
class FlagRerankerCustom:
def __init__(
self,
model_name_or_path: str = None,
use_fp16: bool = False
) -> None:
self.tokenizer = transformers.AutoTokenizer.from_pretrained(
model_name_or_path,
model_max_length=1024,
padding_side="right",
use_fast=False,
trust_remote_code=True
)
self.tokenizer.pad_token_id = self.tokenizer.eod_id
config = transformers.AutoConfig.from_pretrained(
model_name_or_path,
trust_remote_code=True,
bf16=True,
)
config.use_cache = False
self.model = transformers.AutoModelForCausalLM.from_pretrained(
model_name_or_path,
config=config,
trust_remote_code=True,
)
self.model.linear.bfloat16()
if torch.cuda.is_available():
self.device = torch.device('cuda')
elif torch.backends.mps.is_available():
self.device = torch.device('mps')
else:
self.device = torch.device('cpu')
use_fp16 = False
if use_fp16:
self.model.half()
self.model = self.model.to(self.device)
self.model.eval()
self.num_gpus = torch.cuda.device_count()
if self.num_gpus > 1:
print(f"----------using {self.num_gpus}*GPUs----------")
self.model = torch.nn.DataParallel(self.model)
@torch.no_grad()
def compute_score(self, sentence_pairs: Union[List[Tuple[str, str]], Tuple[str, str]], batch_size: int =128,
max_length: int = 1024) -> List[float]:
if self.num_gpus > 0:
batch_size = batch_size * self.num_gpus
assert isinstance(sentence_pairs, list)
if isinstance(sentence_pairs[0], str):
sentence_pairs = [sentence_pairs]
all_scores = []
for start_index in tqdm(range(0, len(sentence_pairs), batch_size), desc="Compute Scores",
disable=False):
sentences_batch = sentence_pairs[start_index:start_index + batch_size] # [[q,ans],[q, ans]...]
inputs = preprocess(sources=sentences_batch, tokenizer=self.tokenizer,max_len=1024,device=self.device)
scores = self.model(**inputs, return_dict=True).logits.view(-1, ).float()
all_scores.extend(scores.cpu().numpy().tolist())
if len(all_scores) == 1:
return all_scores[0]
return all_scores
if __name__ == "__main__":
model_name_or_path = "360Zhinao-1.8B-Reranking"
model = FlagRerankerCustom(model_name_or_path, use_fp16=False)
inputs=[["What Color Is the Sky","Blue"], ["What Color Is the Sky","Pink"],]
ret = model.compute_score(inputs)
print(ret)
License
The source code of this repository follows the open-source license Apache 2.0. 360Zhinao open-source models support commercial use. If you wish to use these models or continue training them for commercial purposes, please contact us via email (g-zhinao-opensource@360.cn) to apply. For the specific license agreement, please see <<360 Zhinao Open-Source Model License>>.
- Downloads last month
- 128
Space using qihoo360/360Zhinao-1.8B-Reranking 1
Evaluation results
- map on MTEB CMedQAv1test set self-reported86.750
- mrr on MTEB CMedQAv1test set self-reported89.154
- map on MTEB CMedQAv2test set self-reported87.916
- mrr on MTEB CMedQAv2test set self-reported89.989
- map on MTEB MMarcoRerankingself-reported37.288
- mrr on MTEB MMarcoRerankingself-reported36.237
- map on MTEB T2Rerankingself-reported68.552
- mrr on MTEB T2Rerankingself-reported79.628