Reason ModernColBERT

🚀 Reason-ModernColBERT

Reason-ModernColBERT是一個後期交互模型，在reasonir-hq數據集上進行訓練。該模型在BRIGHT基準測試中表現出色，該基準測試旨在評估推理密集型檢索性能。Reason-ModernColBERT超越了所有規模達70億參數的現有模型（其規模是該模型的45倍以上），甚至在Stack Exchange分割數據上，平均比ReasonIR-8B（在相同數據上訓練的80億參數模型）的NDCG@10提高了2.5以上。我們將如此出色的結果歸功於後期交互，詳見評估部分。

🚀 快速開始

由於原文檔未提供快速開始的相關內容，此部分暫不展示。

✨ 主要特性

• 高性能：在BRIGHT基準測試中，超越了所有規模達70億參數的現有模型，甚至在Stack Exchange分割數據上，平均比ReasonIR-8B的NDCG@10提高了2.5以上。
• 後期交互：通過後期交互機制，提升了推理密集型檢索的性能。

📦 安裝指南

首先安裝PyLate庫：

pip install -U pylate

💻 使用示例

基礎用法

索引文檔

from pylate import indexes, models, retrieve

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path=pylate_model_id,
)

# Step 2: Initialize the Voyager index
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
    override=True,  # This overwrites the existing index if any
)

# Step 3: Encode the documents
documents_ids = ["1", "2", "3"]
documents = ["document 1 text", "document 2 text", "document 3 text"]

documents_embeddings = model.encode(
    documents,
    batch_size=32,
    is_query=False,  # Ensure that it is set to False to indicate that these are documents, not queries
    show_progress_bar=True,
)

# Step 4: Add document embeddings to the index by providing embeddings and corresponding ids
index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

注意，你不必每次都重新創建索引和編碼文檔。一旦你創建了一個索引並添加了文檔，你可以通過加載它來重複使用該索引：

# To load an index, simply instantiate it with the correct folder/name and without overriding it
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
)

檢索前k個文檔

# Step 1: Initialize the ColBERT retriever
retriever = retrieve.ColBERT(index=index)

# Step 2: Encode the queries
queries_embeddings = model.encode(
    ["query for document 3", "query for document 1"],
    batch_size=32,
    is_query=True,  #  # Ensure that it is set to False to indicate that these are queries
    show_progress_bar=True,
)

# Step 3: Retrieve top-k documents
scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=10,  # Retrieve the top 10 matches for each query
)

高級用法

重排序

如果你只想使用ColBERT模型在你的第一階段檢索管道之上進行重排序，而不構建索引，你可以簡單地使用rank函數並傳入要重排序的查詢和文檔：

from pylate import rank, models

queries = [
    "query A",
    "query B",
]

documents = [
    ["document A", "document B"],
    ["document 1", "document C", "document B"],
]

documents_ids = [
    [1, 2],
    [1, 3, 2],
]

model = models.ColBERT(
    model_name_or_path=pylate_model_id,
)

queries_embeddings = model.encode(
    queries,
    is_query=True,
)

documents_embeddings = model.encode(
    documents,
    is_query=False,
)

reranked_documents = rank.rerank(
    documents_ids=documents_ids,
    queries_embeddings=queries_embeddings,
    documents_embeddings=documents_embeddings,
)

📚 詳細文檔

模型詳情

模型描述

模型來源

• 文檔：PyLate文檔
• 倉庫：GitHub上的PyLate
• Hugging Face：Hugging Face上的PyLate模型

完整模型架構

ColBERT(
  (0): Transformer({'max_seq_length': 127, 'do_lower_case': False}) with Transformer model: ModernBertModel 
  (1): Dense({'in_features': 768, 'out_features': 128, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
)

訓練詳情

訓練數據集

reasonir-hq

• 數據集：train at 0275f82
• 大小：100,521個訓練樣本
• 列：query、pos和neg
• 基於前1000個樣本的近似統計信息： | | 查詢 | 正樣本 | 負樣本 | | ---- | ---- | ---- | ---- | | 類型 | 字符串 | 字符串 | 字符串 | | 詳情 |
  • 最小：38個標記
  • 平均：97.84個標記
  • 最大：128個標記
  |
  • 最小：85個標記
  • 平均：127.63個標記
  • 最大：128個標記
  |
  • 最小：81個標記
  • 平均：127.77個標記
  • 最大：128個標記
  |
• 樣本： | 查詢 | 正樣本 | 負樣本 | | ---- | ---- | ---- | | Given this reasoning-intensive query, find relevant documents that could help answer the question. A researcher is analyzing a sound signal represented by the equation f(t) = 2sin(3πt) + sin(5πt) + 0.5sin(7πt). Using the Fourier transform, what are the frequencies, amplitudes, and phases of the individual sinusoidal components in the signal? | A sound signal is given by the equation f(t) = sin(2πt) + sin(4πt) + sin(6πt) where t is time in seconds. Use Fourier transform to find the frequencies, amplitudes, and phases of the individual sinusoidal components in the signal.
To find the frequencies, amplitudes, and phases of the individual sinusoidal components in the signal f(t) = sin(2πt) + sin(4πt) + sin(6πt), we can use the Fourier transform. The Fourier transform of a continuous function f(t) is given by:

F(ω) = ∫[f(t) * e^(-jωt)] dt

where F(ω) is the Fourier transform of f(t), ω is the angular frequency, and j is the imaginary unit (j^2 = -1). In this case, f(t) is already given as a sum of sinusoidal functions, so we can directly identify the frequencies, amplitudes, and phases of the individual components.

1. First component: sin(2πt)
- Frequency: The angular frequency is 2π, so the frequency is ω/(2π) = 1 Hz.
- Amplitude: The coefficient of the sine function is 1, so the amplitude is 1.
- Phase: There is no phase shi... | The Fourier transform is widely used in various fields, including engineering, physics, and data analysis. It is a powerful tool for decomposing a signal into its constituent frequencies. In music, for example, the Fourier transform can be used to analyze the frequency components of a sound wave. By applying the Fourier transform to a sound signal, one can identify the different frequencies present in the signal, as well as their relative amplitudes. This information can be useful in a variety of applications, such as sound filtering and audio processing. The Fourier transform can also be used to analyze images and other types of data. In image processing, the Fourier transform can be used to filter out noise and other unwanted features from an image. It can also be used to compress images by representing them in the frequency domain. In addition to its many practical applications, the Fourier transform also has a number of interesting theoretical properties. For example, it has been ... | | Given this reasoning-intensive query, find relevant documents that could help answer the question. A manufacturer is designing a cone-shaped container with a fixed volume of 200π cubic centimeters. The container's height is 12 centimeters, and the radius of the base is unknown. If the manufacturer wants to minimize the surface area of the container while maintaining its volume, what should be the radius of the base? | A right circular cone has a radius of 6cm and a slant height of 10cm. Determine the surface area of the cone.
To find the surface area of a right circular cone, we need to calculate the area of the base and the lateral surface area, and then add them together.

The base of the cone is a circle with radius r = 6 cm. The area of the base (A_base) can be found using the formula for the area of a circle:

A_base = πr^2
A_base = π(6 cm)^2
A_base = 36π cm^2

The lateral surface area (A_lateral) can be found using the formula for the lateral surface area of a cone:

A_lateral = πrs, where r is the radius and s is the slant height.

Given that the slant height s = 10 cm, we can calculate the lateral surface area:

A_lateral = π(6 cm)(10 cm)
A_lateral = 60π cm^2

Now, we can find the total surface area (A_total) by adding the base area and the lateral surface area:

A_total = A_base + A_lateral
A_total = 36π cm^2 + 60π cm^2
A_total = 96π cm^2

The surface area of the cone is 96π cm^2. | Torus-Shaped Containers in Chemical Engineering - New Designs and ApplicationsTorus-shaped containers are commonly used in chemical engineering for storing and transporting fluids. These containers have a distinctive doughnut shape, with a central hole and a circular cross-section. In this article, we will explore the design and applications of torus-shaped containers in chemical engineering.One of the main advantages of torus-shaped containers is their high volume-to-surface-area ratio. This makes them ideal for storing large quantities of fluids while minimizing the amount of material needed for construction. Additionally, the curved shape of the container provides added strength and stability, making it less prone to rupture or leakage.The design of torus-shaped containers typically involves the use of computer-aided design (CAD) software to create detailed models of the container's geometry. Engineers can then use these models to simulate various scenarios, such as fluid flow and ... | | Given this reasoning-intensive query, find relevant documents that could help answer the question. On the xy-coordinate plane, points A and B are given as A(2, 4) and B(8, -3). Determine the coordinates of the point on line segment AB that is three times as far from A as it is from B. | On the xy co-ordinate plane, point C is (5,-2) and point D is (-1,1.5). The point on line segment CD that is twice as far from C as from D is:
Answer Choices: (A) (1,-1) (B) (1,1) (C) (2,0.25) (D) (3,0.5) (E) (3,1)
Let's think about the multi-choice question step by step.
We want the point on the line that is twice as far from C as it is from D. We can examine the x and y coordinates separately since they are independent.
*It should be noted that there are two solutions to this problem, one point between C and D, and another point with D in the middle of C and the point. We can quickly look at the answer choices and see that all the points are between C and D, therefore we can search for that point using the following method:
Taking the x-coordinate first, the distance between C and D is |(x-coordinate ofC - (x-coordinate ofD|= |5 - (-1)| = 6
The x-coordinate that is twice as far from C as it is from D (and in between C andD will be 4 units from C and 2 units from D. So the ... | The concept of midpoint is often useful in various mathematical problems, but sometimes we need to find other points that divide a line segment in a particular ratio. One common scenario is when we need to find the point that divides the line segment in the ratio of the other two points. Let's consider an example to understand this better. Suppose we have two points E(3, 4) and F(7, -2) on the xy-coordinate plane, and we want to find the point G on the line segment EF such that EG:GF = 2:5. To solve this problem, we can use the concept of section formula, which states that if a point P(x, y) divides the line segment joining the points A(x1, y1) and B(x2, y2) in the ratio m:n, then the coordinates of P are ((mx2+nx1)/(m+n), (my2+ny1)/(m+n)). Using this formula, we can find the coordinates of point G. First, we need to find the difference in x-coordinates and y-coordinates of points E and F. The difference in x-coordinates is 7 - 3 = 4, and the difference in y-coordinates is -2 - 4 = -6... |
• 損失函數：pylate.losses.cached_contrastive.CachedContrastive

訓練超參數

非默認超參數

• per_device_train_batch_size：256
• per_device_eval_batch_size：256
• learning_rate：1e-05
• bf16：True
• dataloader_num_workers：8

所有超參數

訓練日誌