unstructuredio/SCORE-Bench · Datasets at Hugging Face

text stringlengths 0 2.18k
The information required by Item 9(e) of Schedule 14A will be included under the heading “Disclosure of Fees Paid to Independent Registered Public Accounting Firm” in our definitive proxy statement for our 2024 Annual Meeting of Shareholders, and such required information is incorporated herein by reference.
--------------------------------------------------- Unstructured Page Footer Begin
--------------------------------------------------- Unstructured Page Number Block Begin
70
--------------------------------------------------- Unstructured Page Number Block End
--------------------------------------------------- Unstructured Page Footer End
--------------------------------------------------- Unstructured Plain Text Format 1.0.3

--------------------------------------------------- Unstructured Table Begin
UNSTRUCTURED Compare Product Options
Platform Open Source
For Production AI Workloads For RAG Prototyping
Deployment Options
Serverless SaaS ✓ x
VPC Deployment ✓ ✓
Security and Compliance
One-Click SSO Authentication ✓ x
SOC 2 Type 2 Compliant ✓ x
Connectivity
Supports 25+ source & destination connectors ✓ ✓
Supports 25+ file types ✓ ✓
Supports 30+ types of metadata ✓ ✓
Native integrations with top LLM providers (OpenAI, Anthropic, Octo AI, AWS Bedrock, Voyage AI, Together AI, Databricks, and more) ✓ x
Document Processing
Transforms documents into canonical JSON ✓ ✓
Smart routing to efficiently leverage in-house and third-party models to render any unstructured data RAG-ready ✓ x
Advanced semantic chunking, summary generation, and structured data generation ✓ x
Automatic embedding and chunking algorithm selection ✓ x
Performance
Industry-leading file transformation and table extraction algorithms ✓ x
50x transformation speed with multi-node serving and auto-scaling ✓ x
Usage and Billing
Usage-based pricing with real-time billing and usage dashboard ✓ x
User Interface and API ✓ x
Use via no-code UI or API ✓ x
Additional Features
Workflow management and scheduling ✓ x
Upgraded weekly ✓ x
--------------------------------------------------- Unstructured Table End
--------------------------------------------------- Unstructured Page Footer Begin
UNSTRUCTURED.IO
--------------------------------------------------- Unstructured Page Footer End
--------------------------------------------------- Unstructured Plain Text Format 1
--------------------------------------------------- Unstructured Image Begin
MHSA FFN Pruned Static Skip Dynamic Skip Dynamic
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
Q K V Q K V Q K V Q K V Q K V Q K V
--------------------------------------------------- Unstructured Image End


--------------------------------------------------- Unstructured Caption Begin
Figure 3: An illustration of the optimization results from three types of compression techniques. When reaching the same compression ratio, joint compression will automatically choose optimal dynamic and static percentile schedules to balance the performance and compression ratios.
--------------------------------------------------- Unstructured Caption End


the cross-entropy loss L cls of the image classification task and compression resource loss Lᵣₑₛ. We design the resource loss Lᵣₑₛ as the square of the FLOPs difference between the compressing model and source backbone model. As shown in Eq. 9, F(𝛼; g) represents the FLOPs of the whole ViT model during the optimization of USDC.


--------------------------------------------------- Unstructured Formula Begin
L total = L cls + y L res
L r e s = (F(𝛼;g) — f t ) 2
--------------------------------------------------- Unstructured Formula End


(9)
where fₜ is the target compression ratio and γ is the hyperparameter to balance task loss and resource loss, the value of F (𝛼;g) is affected by static compression parameters 𝛼 and dynamic compression gates g as below:


--------------------------------------------------- Unstructured Formula Begin
F′ attn = 𝛼a l,0 ∑ h∈H 𝛼h l,i,0 F attn
F′ ffn = 𝛼f l,0 ∑ h∈H 𝛼n l,i,1 F ffn
F ( 𝛼;g ) = ∑ l∈L (g l,0 F′ attn +g l,1 F′ ffn + ∑ k∈K 𝛼k l,k F G,l,k ) +F o
--------------------------------------------------- Unstructured Formula End




(10)
Where F′ₐₜₜₙ and F′ ffn are the FLOPs scales of MHSA and FFN block which are normalized by total FLOPs of the model, Fₒ is the FLOPs scale of both embedding layers and final classifier layer in ViT model. F G,l,: are the FLOPs scale of dynamic gates network space of lₜₕ encoder layer.
In the second stage, according to the optimizing results of the static compression parameters 𝛼 on the first stage, we explicitly prune the ViT model and continue to fine-tune the pruned model under dynamic compression constraint. In the second stage, each transformer encoder layer selects a fixed Gₗ with the largest value in


--------------------------------------------------- Unstructured Formula Begin
𝛼g l
--------------------------------------------------- Unstructured Formula End
. The FLOPs F(g) of ViT model on second stage is:


--------------------------------------------------- Unstructured Formula Begin
F (g) = ∑ l∈L (gl,0 F′′ attn +g l,1 F′′ ffn + F G,l ) + F o

The information required by Item 9(e) of Schedule 14A will be included under the heading “Disclosure of Fees Paid to Independent Registered Public Accounting Firm” in our definitive proxy statement for our 2024 Annual Meeting of Shareholders, and such required information is incorporated herein by reference.

--------------------------------------------------- Unstructured Page Footer Begin

--------------------------------------------------- Unstructured Page Number Block Begin

70

--------------------------------------------------- Unstructured Page Number Block End

--------------------------------------------------- Unstructured Page Footer End

--------------------------------------------------- Unstructured Plain Text Format 1.0.3

--------------------------------------------------- Unstructured Table Begin

UNSTRUCTURED Compare Product Options

Platform Open Source

For Production AI Workloads For RAG Prototyping

Deployment Options

Serverless SaaS ✓ x

VPC Deployment ✓ ✓

Security and Compliance

One-Click SSO Authentication ✓ x

SOC 2 Type 2 Compliant ✓ x

Connectivity

Supports 25+ source & destination connectors ✓ ✓

Supports 25+ file types ✓ ✓

Supports 30+ types of metadata ✓ ✓

Native integrations with top LLM providers (OpenAI, Anthropic, Octo AI, AWS Bedrock, Voyage AI, Together AI, Databricks, and more) ✓ x

Document Processing

Transforms documents into canonical JSON ✓ ✓

Smart routing to efficiently leverage in-house and third-party models to render any unstructured data RAG-ready ✓ x

Advanced semantic chunking, summary generation, and structured data generation ✓ x

Automatic embedding and chunking algorithm selection ✓ x

Performance

Industry-leading file transformation and table extraction algorithms ✓ x

50x transformation speed with multi-node serving and auto-scaling ✓ x

Usage and Billing

Usage-based pricing with real-time billing and usage dashboard ✓ x

User Interface and API ✓ x

Use via no-code UI or API ✓ x

Additional Features

Workflow management and scheduling ✓ x

Upgraded weekly ✓ x

--------------------------------------------------- Unstructured Table End

--------------------------------------------------- Unstructured Page Footer Begin

UNSTRUCTURED.IO

--------------------------------------------------- Unstructured Page Footer End

--------------------------------------------------- Unstructured Plain Text Format 1

--------------------------------------------------- Unstructured Image Begin

MHSA FFN Pruned Static Skip Dynamic Skip Dynamic

Q K V Q K V Q K V Q K V Q K V Q K V

--------------------------------------------------- Unstructured Image End

--------------------------------------------------- Unstructured Caption Begin

Figure 3: An illustration of the optimization results from three types of compression techniques. When reaching the same compression ratio, joint compression will automatically choose optimal dynamic and static percentile schedules to balance the performance and compression ratios.

--------------------------------------------------- Unstructured Caption End

the cross-entropy loss L cls of the image classification task and compression resource loss Lᵣₑₛ. We design the resource loss Lᵣₑₛ as the square of the FLOPs difference between the compressing model and source backbone model. As shown in Eq. 9, F(𝛼; g) represents the FLOPs of the whole ViT model during the optimization of USDC.

--------------------------------------------------- Unstructured Formula Begin

L total = L cls + y L res

L r e s = (F(𝛼;g) — f t ) 2

--------------------------------------------------- Unstructured Formula End

(9)

where fₜ is the target compression ratio and γ is the hyperparameter to balance task loss and resource loss, the value of F (𝛼;g) is affected by static compression parameters 𝛼 and dynamic compression gates g as below:

--------------------------------------------------- Unstructured Formula Begin

F′ attn = 𝛼a l,0 ∑ h∈H 𝛼h l,i,0 F attn

F′ ffn = 𝛼f l,0 ∑ h∈H 𝛼n l,i,1 F ffn

F ( 𝛼;g ) = ∑ l∈L (g l,0 F′ attn +g l,1 F′ ffn + ∑ k∈K 𝛼k l,k F G,l,k ) +F o

--------------------------------------------------- Unstructured Formula End

(10)

Where F′ₐₜₜₙ and F′ ffn are the FLOPs scales of MHSA and FFN block which are normalized by total FLOPs of the model, Fₒ is the FLOPs scale of both embedding layers and final classifier layer in ViT model. F G,l,: are the FLOPs scale of dynamic gates network space of lₜₕ encoder layer.

In the second stage, according to the optimizing results of the static compression parameters 𝛼 on the first stage, we explicitly prune the ViT model and continue to fine-tune the pruned model under dynamic compression constraint. In the second stage, each transformer encoder layer selects a fixed Gₗ with the largest value in

--------------------------------------------------- Unstructured Formula Begin

𝛼g l

--------------------------------------------------- Unstructured Formula End

. The FLOPs F(g) of ViT model on second stage is:

--------------------------------------------------- Unstructured Formula Begin

F (g) = ∑ l∈L (gl,0 F′′ attn +g l,1 F′′ ffn + F G,l ) + F o