Granite-Embedding-30m-English (修订版 r1.1)

模型概述： Granite-Embedding-30m-English 是 Granite Embeddings 系列中一款拥有 3000 万参数的密集双编码器嵌入模型，可用于生成高质量的文本嵌入。该模型生成的嵌入向量维度为 384，采用开源相关性配对数据集（具有宽松的企业友好许可证）和 IBM 收集及生成的数据集进行训练。在保持 BEIR 等学术基准测试竞争力的同时，该模型在许多企业应用场景中也有良好表现。该模型采用检索导向的预训练、对比微调、知识蒸馏和模型合并等技术开发，以提高性能。

Granite-embedding-30m-r1.1 专门为支持多轮信息检索而设计，用于处理多轮会话信息检索中的上下文文档检索。Granite-embedding-30m-r1.1 基于多轮会话信息检索定制的数据进行训练，并使用多教师蒸馏技术对 granite-embedding-30m-english (https://huggingface.co/ibm-granite/granite-embedding-30m-english) 进行蒸馏。

• 开发者： Granite Embedding Team, IBM
• GitHub 仓库： ibm-granite/granite-embedding-models
• 网站： Granite Docs
• 论文： Technical Report
• 发布日期： 2025年8月29日（granite-embedding-30m-english-r1.1）
• 许可证： Apache 2.0

支持的语言： 英语。

预期用途： 该模型旨在为给定文本生成固定长度的向量表示，可用于文本相似度、检索和搜索应用。

与 Sentence Transformers 配合使用： 该模型与 SentenceTransformer 库兼容，使用非常简单：

首先，安装 sentence transformers 库

pip install sentence_transformers

然后可以使用该模型对文本对进行编码，并找出它们表示之间的相似度。

Granite-Embedding-30m-English

from sentence_transformers import SentenceTransformer, util

model_path = "ibm-granite/granite-embedding-30m-english"
# 加载 Sentence Transformer 模型
model = SentenceTransformer(model_path)

input_queries = [
    ' Who made the song My achy breaky heart? ',
    'summit define'
    ]

input_passages = [
    "Achy Breaky Heart is a country song written by Don Von Tress. Originally titled Don't Tell My Heart and performed by The Marcy Brothers in 1991. ",
    "Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments."
    ]

# 对查询和段落进行编码
query_embeddings = model.encode(input_queries)
passage_embeddings = model.encode(input_passages)

# 计算余弦相似度
print(util.cos_sim(query_embeddings, passage_embeddings))

Granite-Embedding-30m-r1.1

具体而言，要使用 granite-embedding-30m-r1.1 对整个对话进行编码，需要将对话的最后用户查询作为输入，并按时间倒序排列对话实例：首先是最后的用户查询，其次是之前的代理响应，然后是更早的用户查询。例如：

对话：user:<user_query_1> agent: <agent_response_1> user:<user_query_2> agent: <agent_response_2> user:<user_query_3> agent: <agent_response_3> user:<last_user_query>

转换为输入查询格式：<last_user_query>[SEP]agent: <agent_response_3>||user:<user_query_3>||agent: <agent_response_2>||user:<user_query_2>||agent: <agent_response_1>||user:<user_query_1>

from sentence_transformers import SentenceTransformer, util

model_path = "ibm-granite/granite-embedding-30m-english"
# 加载 Sentence Transformer 模型
model = SentenceTransformer(model_path, revision="granite-embedding-30m-r1.1")

input_queries = [
    "Which team has won the most Super Bowls?[SEP]agent: Six teams from each conference (AFC and NFC), for a total of 12 team playoff system.||user: How many teams are in the NFL playoffs?||agent: There are 32 teams in the National Football League (NFL).||user: How many teams are in the NFL?",

    "How many teams are in the NFL playoffs?[SEP]agent: There are 32 teams in the National Football League (NFL).||user: How many teams are in the NFL?||agent: The Chicago Cardinals became the St. Louis Cardinals in 1960 and eventually moved and became the Arizona Cardinals. The Chicago Cardinals ( now the Arizona Cardinals ) were a founding member of the NFL.||user: Are the Arizona Cardinals and the Chicago Cardinals the same team?||agent: The Arizona Cardinals do play outside the United States. They had a game in London, England, on October 22, 2017, against the Los Angeles Rams at Twickenham Stadium and in 2005 they played in Mexico.||user: Do the Arizona Cardinals play outside the US?"
    ]

input_passages = [
    "Super Bowl\nThe Pittsburgh Steelers have won six Super Bowls , the most of any team ; the Dallas Cowboys , New England Patriots and San Francisco 49ers have five victories each , while the Green Bay Packers and New York Giants have four Super Bowl championships . Fourteen other NFL franchises have won at least one Super Bowl . Eight teams have appeared in Super Bowl games without a win . The Minnesota Vikings were the first team to have appeared a record four times without a win . The Buffalo Bills played in a record four Super Bowls in a row and lost every one . Four teams ( the Cleveland Browns , Detroit Lions , Jacksonville Jaguars , and Houston Texans ) have never appeared in a Super Bowl . The Browns and Lions both won NFL Championships prior to the creation of the Super Bowl , while the Jaguars ( 1995 ) and Texans ( 2002 ) are both recent NFL expansion teams . ( Detroit , Houston , and Jacksonville , however , have hosted a Super Bowl , leaving the Browns the only team to date who has neither played in nor whose city has hosted the game . ) The Minnesota Vikings won the last NFL Championship before the merger but lost to the AFL champion Kansas City Chiefs in Super Bowl IV.",

    "NFL playoffs \n The 32 - team National Football League is divided into two conferences , American Football Conference ( AFC ) and National Football Conference ( NFC ) , each with 16 teams . Since 2002 , each conference has been further divided into four divisions of four teams each . The tournament brackets are made up of six teams from each of the league 's two conferences , following the end of the regular season . Qualification into the playoffs works as follows : "
    ]

# 对查询和段落进行编码
query_embeddings = model.encode(input_queries)
passage_embeddings = model.encode(input_passages)

# 计算余弦相似度
print(util.cos_sim(query_embeddings, passage_embeddings))

与 Huggingface Transformers 配合使用： 以下是一个简单的示例，展示如何使用 Transformers 库和 PyTorch 来使用 Granite-Embedding-30m-English 模型。

首先，安装所需的库

pip install transformers torch

然后可以使用该模型对文本对进行编码

Granite-Embedding-30m-English

import torch
from transformers import AutoModel, AutoTokenizer

model_path = "ibm-granite/granite-embedding-30m-english"

# 加载模型和分词器
model = AutoModel.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model.eval()

input_queries = [
    ' Who made the song My achy breaky heart? ',
    'summit define'
    ]

# 对输入进行分词
tokenized_queries = tokenizer(input_queries, padding=True, truncation=True, return_tensors='pt')

# 对查询进行编码
with torch.no_grad():
    # 查询
    model_output = model(**tokenized_queries)
    # 执行池化。granite-embedding-30m-english 使用 CLS 池化
    query_embeddings = model_output[0][:, 0]

# 对嵌入进行归一化
query_embeddings = torch.nn.functional.normalize(query_embeddings, p=2, dim=1)

Granite-Embedding-30m-r1.1

import torch
from transformers import AutoModel, AutoTokenizer

model_path = "ibm-granite/granite-embedding-30m-english"

# 加载模型和分词器
model = AutoModel.from_pretrained(model_path, revision="granite-embedding-30m-r1.1")
tokenizer = AutoTokenizer.from_pretrained(model_path)
model.eval()

input_queries = [
    "Which team has won the most Super Bowls?[SEP]agent: Six teams from each conference (AFC and NFC), for a total of 12 team playoff system.||user: How many teams are in the NFL playoffs?||agent: There are 32 teams in the National Football League (NFL).||user: How many teams are in the NFL?",

    "How many teams are in the NFL playoffs?[SEP]agent: There are 32 teams in the National Football League (NFL).||user: How many teams are in the NFL?||agent: The Chicago Cardinals became the St. Louis Cardinals in 1960 and eventually moved and became the Arizona Cardinals. The Chicago Cardinals ( now the Arizona Cardinals ) were a founding member of the NFL.||user: Are the Arizona Cardinals and the Chicago Cardinals the same team?||agent: The Arizona Cardinals do play outside the United States. They had a game in London, England, on October 22, 2017, against the Los Angeles Rams at Twickenham Stadium and in 2005 they played in Mexico.||user: Do the Arizona Cardinals play outside the US?"
    ]

# 对输入进行分词
tokenized_queries = tokenizer(input_queries, padding=True, truncation=True, return_tensors='pt')

# 对查询进行编码
with torch.no_grad():
    # 查询
    model_output = model(**tokenized_queries)
    # 执行池化。granite-embedding-30m-english-multiturn 使用 CLS 池化
    query_embeddings = model_output[0][:, 0]

# 对嵌入进行归一化
query_embeddings = torch.nn.functional.normalize(query_embeddings, p=2, dim=1)

评估：

Granite-Embedding-30M-English 模型比其他具有相似嵌入维度的模型快一倍，同时保持竞争力的性能。以下是 Granite-Embedding-30M-English 模型在 MTEB Retrieval（即 BEIR）和代码检索（CoIR）基准测试上的表现。

granite-embedding-30m-r1.1 修订版保持了 granite-embedding-30m-english 的快速速度，同时在多轮信息检索基准测试上表现出色。以下是 granite-embedding-30M-r1.1 模型在 MTEB Retrieval（即 BEIR）和多轮信息检索（MTRAG(https://github.com/IBM/mt-rag-benchmark)、Multidoc2dial(https://github.com/IBM/multidoc2dial)）数据集上的表现。

模型架构： granite-embedding-30m-English 基于类似 RoBERTa 的纯编码器 Transformer 架构，在 IBM Research 内部训练。granite-embedding-30m-r1.1 与 granite-embedding-30m-English 采用相同的架构。

训练数据： 总体而言，训练数据主要来自四个关键来源：（1）从网页抓取的无监督标题-正文配对数据，（2）具有宽松、企业友好许可证的公开配对数据，（3）针对特定技术领域的 IBM 内部配对数据，以及（4）IBM 生成的合成数据。数据详情如下：

值得注意的是，我们没有在训练语料库中使用流行的 MS-MARCO 检索数据集，因为它采用非商业许可证，而其他开源模型由于其高质量而使用该数据集进行训练。

基础设施： 我们使用 IBM 的计算集群 Cognitive Compute Cluster 来训练 Granite Embedding 模型，该集群配备 NVIDIA A100 80gb GPU。该集群为我们提供了可扩展且高效的基础设施，支持在多个 GPU 上训练模型。

伦理考量和局限性： 用于训练基础语言模型的数据经过过滤，移除了包含仇恨、虐待和脏话的文本。Granite-embedding-30m-english 和 Granite-embedding-30m-r1.1 仅针对英语文本进行训练，上下文长度为 512 个标记（超过此长度的文本将被截断）。

资源

• ⭐️ 了解 Granite 的最新更新：https://www.ibm.com/granite
• 📄 开始使用教程、最佳实践和提示工程建议：https://www.ibm.com/granite/docs/
• 💡 了解最新的 Granite 学习资源：https://ibm.biz/granite-learning-resources