Reranker Bert Tiny Gooaq Bce Tanh V3
这是一个基于BERT-tiny微调的交叉编码器模型,用于计算文本对的相似度分数,适用于语义搜索、文本分类等任务。
下载量 1,962
发布时间 : 3/4/2025
模型简介
该模型通过sentence-transformers库开发,能够计算文本对的相似度分数,可用于语义文本相似度、语义搜索、复述挖掘、文本分类、聚类等任务。
模型特点
高效轻量
基于BERT-tiny架构,模型体积小,推理速度快
语义相关性评估
能够准确评估文本对之间的语义相关性
大规模训练
在578,402条GooAQ数据上进行训练
模型能力
文本相似度计算
语义搜索重排序
问答对匹配
文本分类
使用案例
信息检索
搜索引擎结果重排序
对搜索引擎返回的结果进行相关性重排序
在gooaq-dev数据集上map达到0.5677
问答系统
问答对匹配
评估问题与候选答案的相关性
🚀 BERT-tiny在GooAQ上训练的模型
这是一个基于Cross Encoder的模型,它使用sentence-transformers库从prajjwal1/bert-tiny微调而来。该模型可以为文本对计算得分,可用于语义文本相似度、语义搜索、释义挖掘、文本分类、聚类等任务。
此模型使用train_script.py进行训练。
🚀 快速开始
本模型是一个基于Cross Encoder
的微调模型,可用于计算文本对的得分,适用于多种自然语言处理任务。下面将介绍如何安装依赖库并使用该模型进行推理。
✨ 主要特性
- 多任务适用性:可用于语义文本相似度、语义搜索、释义挖掘、文本分类、聚类等多种任务。
- 微调模型:基于
prajjwal1/bert-tiny
进行微调,能更好地适应特定任务。
📦 安装指南
首先,你需要安装Sentence Transformers
库:
pip install -U sentence-transformers
💻 使用示例
基础用法
安装好库后,你可以加载模型并进行推理:
from sentence_transformers import CrossEncoder
# 从🤗 Hub下载模型
model = CrossEncoder("cross-encoder-testing/reranker-bert-tiny-gooaq-bce")
# 定义文本对
pairs = [
['are javascript developers in demand?', "JavaScript is the skill that is most in-demand for IT in 2020, according to a report from developer skills tester DevSkiller. The report, “Top IT Skills report 2020: Demand and Hiring Trends,” has JavaScript switching places with Java when compared to last year's report, with Java in third place this year, behind SQL."],
['are javascript developers in demand?', 'In one line difference between the two is: JavaScript is the programming language where as AngularJS is a framework based on JavaScript. ... It is also the basic for all java script based technologies like jquery, angular JS, bootstrap JS and so on. Angular JS is a framework written in javascript and uses MVC architecture.'],
['are javascript developers in demand?', 'Java applications are run in a virtual machine or web browser while JavaScript is run on a web browser. Java code is compiled whereas while JavaScript code is in text and in a web page. JavaScript is an OOP scripting language, whereas Java is an OOP programming language.'],
['are javascript developers in demand?', 'Things in the body tag are the things that should be displayed: the actual content. Javascript in the body is executed as it is read and as the page is rendered. Javascript in the head is interpreted before anything is rendered.'],
['are javascript developers in demand?', 'Web apps tend to be built using JavaScript, CSS and HTML5. Unlike mobile apps, there is no standard software development kit for building web apps. However, developers do have access to templates. Compared to mobile apps, web apps are usually quicker and easier to build — but they are much simpler in terms of features.'],
]
# 预测得分
scores = model.predict(pairs)
print(scores.shape)
# (5,)
# 或者根据与单个文本的相似度对不同文本进行排序
ranks = model.rank(
'are javascript developers in demand?',
[
"JavaScript is the skill that is most in-demand for IT in 2020, according to a report from developer skills tester DevSkiller. The report, “Top IT Skills report 2020: Demand and Hiring Trends,” has JavaScript switching places with Java when compared to last year's report, with Java in third place this year, behind SQL.",
'In one line difference between the two is: JavaScript is the programming language where as AngularJS is a framework based on JavaScript. ... It is also the basic for all java script based technologies like jquery, angular JS, bootstrap JS and so on. Angular JS is a framework written in javascript and uses MVC architecture.',
'Java applications are run in a virtual machine or web browser while JavaScript is run on a web browser. Java code is compiled whereas while JavaScript code is in text and in a web page. JavaScript is an OOP scripting language, whereas Java is an OOP programming language.',
'Things in the body tag are the things that should be displayed: the actual content. Javascript in the body is executed as it is read and as the page is rendered. Javascript in the head is interpreted before anything is rendered.',
'Web apps tend to be built using JavaScript, CSS and HTML5. Unlike mobile apps, there is no standard software development kit for building web apps. However, developers do have access to templates. Compared to mobile apps, web apps are usually quicker and easier to build — but they are much simpler in terms of features.',
]
)
# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]
📚 详细文档
模型详情
属性 | 详情 |
---|---|
模型类型 | Cross Encoder |
基础模型 | prajjwal1/bert-tiny |
最大序列长度 | 512 tokens |
输出标签数量 | 1 label |
语言 | en |
许可证 | apache-2.0 |
模型资源
- 文档:Sentence Transformers Documentation
- 文档:Cross Encoder Documentation
- 仓库:Sentence Transformers on GitHub
- Hugging Face:Cross Encoders on Hugging Face
评估指标
Cross Encoder重排序
- 数据集:
gooaq-dev
、NanoMSMARCO
、NanoNFCorpus
和NanoNQ
- 评估方法:使用
CrossEncoderRerankingEvaluator
进行评估
指标 | gooaq-dev | NanoMSMARCO | NanoNFCorpus | NanoNQ |
---|---|---|---|---|
map | 0.5677 (+0.0366) | 0.4280 (-0.0616) | 0.3397 (+0.0787) | 0.4149 (-0.0047) |
mrr@10 | 0.5558 (+0.0318) | 0.4129 (-0.0646) | 0.5196 (+0.0198) | 0.4132 (-0.0135) |
ndcg@10 | 0.6157 (+0.0245) | 0.4772 (-0.0632) | 0.3308 (+0.0058) | 0.4859 (-0.0147) |
Cross Encoder Nano BEIR
- 数据集:
NanoBEIR_R100_mean
- 评估方法:使用
CrossEncoderNanoBEIREvaluator
进行评估
指标 | 值 |
---|---|
map | 0.3942 (+0.0041) |
mrr@10 | 0.4486 (-0.0194) |
ndcg@10 | 0.4313 (-0.0241) |
训练详情
训练数据集
- 未命名数据集
- 大小:578,402个训练样本
- 列:
question
、answer
和label
- 基于前1000个样本的近似统计信息:
| | 问题 | 答案 | 标签 |
| ---- | ---- | ---- | ---- |
| 类型 | string | string | int |
| 详情 |
- 最小长度: 21个字符
- 平均长度: 43.81个字符
- 最大长度: 96个字符
- 最小长度: 51个字符
- 平均长度: 252.46个字符
- 最大长度: 405个字符
- 0: ~82.90%
- 1: ~17.10%
- 样本:
| 问题 | 答案 | 标签 |
| ---- | ---- | ---- |
|
are javascript developers in demand?
|JavaScript is the skill that is most in-demand for IT in 2020, according to a report from developer skills tester DevSkiller. The report, “Top IT Skills report 2020: Demand and Hiring Trends,” has JavaScript switching places with Java when compared to last year's report, with Java in third place this year, behind SQL.
|1
| |are javascript developers in demand?
|In one line difference between the two is: JavaScript is the programming language where as AngularJS is a framework based on JavaScript. ... It is also the basic for all java script based technologies like jquery, angular JS, bootstrap JS and so on. Angular JS is a framework written in javascript and uses MVC architecture.
|0
| |are javascript developers in demand?
|Java applications are run in a virtual machine or web browser while JavaScript is run on a web browser. Java code is compiled whereas while JavaScript code is in text and in a web page. JavaScript is an OOP scripting language, whereas Java is an OOP programming language.
|0
|
- 损失函数:使用
BinaryCrossEntropyLoss
,参数如下:
{
"activation_fct": "torch.nn.modules.linear.Identity",
"pos_weight": 5
}
训练超参数
非默认超参数
eval_strategy
: stepsper_device_train_batch_size
: 2048per_device_eval_batch_size
: 2048learning_rate
: 0.0005num_train_epochs
: 1warmup_ratio
: 0.1seed
: 12bf16
: True
所有超参数
点击展开
overwrite_output_dir
: Falsedo_predict
: Falseeval_strategy
: stepsprediction_loss_only
: Trueper_device_train_batch_size
: 2048per_device_eval_batch_size
: 2048per_gpu_train_batch_size
: Noneper_gpu_eval_batch_size
: Nonegradient_accumulation_steps
: 1eval_accumulation_steps
: Nonetorch_empty_cache_steps
: Nonelearning_rate
: 0.0005weight_decay
: 0.0adam_beta1
: 0.9adam_beta2
: 0.999adam_epsilon
: 1e-08max_grad_norm
: 1.0num_train_epochs
: 1max_steps
: -1lr_scheduler_type
: linearlr_scheduler_kwargs
: {}warmup_ratio
: 0.1warmup_steps
: 0log_level
: passivelog_level_replica
: warninglog_on_each_node
: Truelogging_nan_inf_filter
: Truesave_safetensors
: Truesave_on_each_node
: Falsesave_only_model
: Falserestore_callback_states_from_checkpoint
: Falseno_cuda
: Falseuse_cpu
: Falseuse_mps_device
: Falseseed
: 12data_seed
: Nonejit_mode_eval
: Falseuse_ipex
: Falsebf16
: Truefp16
: Falsefp16_opt_level
: O1half_precision_backend
: autobf16_full_eval
: Falsefp16_full_eval
: Falsetf32
: Nonelocal_rank
: 0ddp_backend
: Nonetpu_num_cores
: Nonetpu_metrics_debug
: Falsedebug
: []dataloader_drop_last
: Falsedataloader_num_workers
: 0dataloader_prefetch_factor
: Nonepast_index
: -1disable_tqdm
: Falseremove_unused_columns
: Truelabel_names
: Noneload_best_model_at_end
: Falseignore_data_skip
: Falsefsdp
: []fsdp_min_num_params
: 0fsdp_config
: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}fsdp_transformer_layer_cls_to_wrap
: Noneaccelerator_config
: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}deepspeed
: Nonelabel_smoothing_factor
: 0.0optim
: adamw_torchoptim_args
: Noneadafactor
: Falsegroup_by_length
: Falselength_column_name
: lengthddp_find_unused_parameters
: Noneddp_bucket_cap_mb
: Noneddp_broadcast_buffers
: Falsedataloader_pin_memory
: Truedataloader_persistent_workers
: Falseskip_memory_metrics
: Trueuse_legacy_prediction_loop
: Falsepush_to_hub
: Falseresume_from_checkpoint
: Nonehub_model_id
: Nonehub_strategy
: every_savehub_private_repo
: Nonehub_always_push
: Falsegradient_checkpointing
: Falsegradient_checkpointing_kwargs
: Noneinclude_inputs_for_metrics
: Falseinclude_for_metrics
: []eval_do_concat_batches
: Truefp16_backend
: autopush_to_hub_model_id
: Nonepush_to_hub_organization
: Nonemp_parameters
:auto_find_batch_size
: Falsefull_determinism
: Falsetorchdynamo
: Noneray_scope
: lastddp_timeout
: 1800torch_compile
: Falsetorch_compile_backend
: Nonetorch_compile_mode
: Nonedispatch_batches
: Nonesplit_batches
: Noneinclude_tokens_per_second
: Falseinclude_num_input_tokens_seen
: Falseneftune_noise_alpha
: Noneoptim_target_modules
: Nonebatch_eval_metrics
: Falseeval_on_start
: Falseuse_liger_kernel
: Falseeval_use_gather_object
: Falseaverage_tokens_across_devices
: Falseprompts
: Nonebatch_sampler
: batch_samplermulti_dataset_batch_sampler
: proportional
训练日志
轮次 | 步数 | 训练损失 | gooaq-dev_ndcg@10 | NanoMSMARCO_ndcg@10 | NanoNFCorpus_ndcg@10 | NanoNQ_ndcg@10 | NanoBEIR_R100_mean_ndcg@10 |
---|---|---|---|---|---|---|---|
-1 | -1 | - | 0.0887 (-0.5025) | 0.0063 (-0.5341) | 0.3262 (+0.0012) | 0.0000 (-0.5006) | 0.1108 (-0.3445) |
0.0035 | 1 | 1.1945 | - | - | - | - | - |
0.0707 | 20 | 1.1664 | 0.4082 (-0.1830) | 0.1805 (-0.3600) | 0.3168 (-0.0083) | 0.2243 (-0.2763) | 0.2405 (-0.2149) |
0.1413 | 40 | 1.1107 | 0.5260 (-0.0652) | 0.3453 (-0.1951) | 0.3335 (+0.0085) | 0.3430 (-0.1576) | 0.3406 (-0.1147) |
0.2120 | 60 | 1.022 | 0.5623 (-0.0289) | 0.3929 (-0.1475) | 0.3512 (+0.0262) | 0.3472 (-0.1535) | 0.3638 (-0.0916) |
0.2827 | 80 | 0.973 | 0.5691 (-0.0221) | 0.4048 (-0.1356) | 0.3530 (+0.0280) | 0.3833 (-0.1174) | 0.3804 (-0.0750) |
0.3534 | 100 | 0.963 | 0.5814 (-0.0098) | 0.4385 (-0.1019) | 0.3471 (+0.0221) | 0.4227 (-0.0779) | 0.4028 (-0.0526) |
0.4240 | 120 | 0.9419 | 0.5963 (+0.0050) | 0.4106 (-0.1298) | 0.3540 (+0.0289) | 0.4843 (-0.0163) | 0.4163 (-0.0391) |
0.4947 | 140 | 0.9331 | 0.5953 (+0.0041) | 0.4310 (-0.1094) | 0.3367 (+0.0117) | 0.4163 (-0.0843) | 0.3947 (-0.0607) |
0.5654 | 160 | 0.9263 | 0.6070 (+0.0158) | 0.4626 (-0.0778) | 0.3443 (+0.0193) | 0.4823 (-0.0184) | 0.4297 (-0.0256) |
0.6360 | 180 | 0.9212 | 0.6069 (+0.0156) | 0.4602 (-0.0802) | 0.3391 (+0.0141) | 0.4782 (-0.0224) | 0.4258 (-0.0295) |
0.7067 | 200 | 0.901 | 0.6126 (+0.0214) | 0.4602 (-0.0803) | 0.3413 (+0.0162) | 0.4780 (-0.0227) | 0.4265 (-0.0289) |
0.7774 | 220 | 0.8997 | 0.6136 (+0.0224) | 0.4801 (-0.0604) | 0.3349 (+0.0098) | 0.4903 (-0.0103) | 0.4351 (-0.0203) |
0.8481 | 240 | 0.9021 | 0.6132 (+0.0220) | 0.4850 (-0.0554) | 0.3438 (+0.0188) | 0.4855 (-0.0151) | 0.4381 (-0.0173) |
0.9187 | 260 | 0.9013 | 0.6188 (+0.0276) | 0.4820 (-0.0584) | 0.3387 (+0.0137) | 0.4851 (-0.0156) | 0.4353 (-0.0201) |
0.9894 | 280 | 0.8996 | 0.6157 (+0.0245) | 0.4772 (-0.0632) | 0.3305 (+0.0054) | 0.4859 (-0.0147) | 0.4312 (-0.0242) |
-1 | -1 | - | 0.6157 (+0.0245) | 0.4772 (-0.0632) | 0.3308 (+0.0058) | 0.4859 (-0.0147) | 0.4313 (-0.0241) |
环境影响
使用CodeCarbon测量碳排放:
- 能耗:0.019 kWh
- 碳排放:0.007 kg的CO2
- 使用时长:0.099小时
训练硬件
- 是否使用云服务:否
- GPU型号:1 x NVIDIA GeForce RTX 3090
- CPU型号:13th Gen Intel(R) Core(TM) i7-13700K
- 内存大小:31.78 GB
框架版本
- Python: 3.11.6
- Sentence Transformers: 3.5.0.dev0
- Transformers: 4.48.3
- PyTorch: 2.5.0+cu121
- Accelerate: 1.3.0
- Datasets: 2.20.0
- Tokenizers: 0.21.0
🔧 技术细节
本模型基于Cross Encoder
架构,使用BinaryCrossEntropyLoss
作为损失函数进行微调。通过对特定数据集的训练,模型能够学习到文本对之间的语义关系,从而为文本对计算得分。在训练过程中,使用了一系列超参数来控制训练过程,如学习率、批次大小等。
📄 许可证
本模型使用apache-2.0
许可证。
📖 引用
BibTeX
Sentence Transformers
@inproceedings{reimers-2019-sentence-bert,
title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
author = "Reimers, Nils and Gurevych, Iryna",
booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
month = "11",
year = "2019",
publisher = "Association for Computational Linguistics",
url = "https://arxiv.org/abs/1908.10084",
}
Jina Embeddings V3
Jina Embeddings V3 是一个多语言句子嵌入模型,支持超过100种语言,专注于句子相似度和特征提取任务。
文本嵌入
Transformers 支持多种语言

J
jinaai
3.7M
911
Ms Marco MiniLM L6 V2
Apache-2.0
基于MS Marco段落排序任务训练的交叉编码器模型,用于信息检索中的查询-段落相关性评分
文本嵌入 英语
M
cross-encoder
2.5M
86
Opensearch Neural Sparse Encoding Doc V2 Distill
Apache-2.0
基于蒸馏技术的稀疏检索模型,专为OpenSearch优化,支持免推理文档编码,在搜索相关性和效率上优于V1版本
文本嵌入
Transformers 英语

O
opensearch-project
1.8M
7
Sapbert From PubMedBERT Fulltext
Apache-2.0
基于PubMedBERT的生物医学实体表征模型,通过自对齐预训练优化语义关系捕捉
文本嵌入 英语
S
cambridgeltl
1.7M
49
Gte Large
MIT
GTE-Large 是一个强大的句子转换器模型,专注于句子相似度和文本嵌入任务,在多个基准测试中表现出色。
文本嵌入 英语
G
thenlper
1.5M
278
Gte Base En V1.5
Apache-2.0
GTE-base-en-v1.5 是一个英文句子转换器模型,专注于句子相似度任务,在多个文本嵌入基准测试中表现优异。
文本嵌入
Transformers 支持多种语言

G
Alibaba-NLP
1.5M
63
Gte Multilingual Base
Apache-2.0
GTE Multilingual Base 是一个多语言的句子嵌入模型,支持超过50种语言,适用于句子相似度计算等任务。
文本嵌入
Transformers 支持多种语言

G
Alibaba-NLP
1.2M
246
Polybert
polyBERT是一个化学语言模型,旨在实现完全由机器驱动的超快聚合物信息学。它将PSMILES字符串映射为600维密集指纹,以数值形式表示聚合物化学结构。
文本嵌入
Transformers

P
kuelumbus
1.0M
5
Bert Base Turkish Cased Mean Nli Stsb Tr
Apache-2.0
基于土耳其语BERT的句子嵌入模型,专为语义相似度任务优化
文本嵌入
Transformers 其他

B
emrecan
1.0M
40
GIST Small Embedding V0
MIT
基于BAAI/bge-small-en-v1.5模型微调的文本嵌入模型,通过MEDI数据集与MTEB分类任务数据集训练,优化了检索任务的查询编码能力。
文本嵌入
Safetensors 英语
G
avsolatorio
945.68k
29
精选推荐AI模型
Llama 3 Typhoon V1.5x 8b Instruct
专为泰语设计的80亿参数指令模型,性能媲美GPT-3.5-turbo,优化了应用场景、检索增强生成、受限生成和推理任务
大型语言模型
Transformers 支持多种语言

L
scb10x
3,269
16
Cadet Tiny
Openrail
Cadet-Tiny是一个基于SODA数据集训练的超小型对话模型,专为边缘设备推理设计,体积仅为Cosmo-3B模型的2%左右。
对话系统
Transformers 英语

C
ToddGoldfarb
2,691
6
Roberta Base Chinese Extractive Qa
基于RoBERTa架构的中文抽取式问答模型,适用于从给定文本中提取答案的任务。
问答系统 中文
R
uer
2,694
98
智启未来,您的人工智能解决方案智库
简体中文