GOTOCR2Options

Options & Configuration · AiDotNet.VisionLanguage.Document

Configuration options for GOT-OCR2, the General OCR Theory model (Wei et al., 2024).

Inheritance: ModelOptions → GenerativeVLMOptions → DocumentVLMOptions → GOTOCR2Options

How It Works

GOT-OCR2 has three parts:

  • A SAM ViTDet-B image encoder over 1024 px pages: patch 16, 768 wide, 12 blocks, windowed 14x14

attention with global attention in blocks 2/5/8/11, and a 256-channel neck.

  • Two stride-2 convolutions that reduce the page to 256 image tokens.
  • The Qwen-0.5B decoder: 1024 wide, 24 layers, 16 heads with 16 KV heads, a SiLU-gated FFN of 2816,

RoPE theta 1e6, and a 151860-token vocabulary.

These are the defaults of HF GotOcr2Config.

The inherited options mean the following here: VisionDim, NumVisionLayers and NumHeads describe the ViTDet encoder. DecoderDim and NumDecoderLayers describe the Qwen decoder. MaxSequenceLength bounds the full prompt plus answer.

Constructors

Constructor Summary
GOTOCR2Options() Creates the published GOT-OCR2 configuration.
GOTOCR2Options(GOTOCR2Options other) Copies every option from other.

Properties

Property Summary
bool FormatOutput { get; set; } When true, the query is “OCR with format: “, which asks for formatted output (Markdown, LaTeX, music notation).
double RmsNormEpsilon { get; set; } RMSNorm epsilon.
double RopeTheta { get; set; } Rotary embedding base.
int DecoderFeedForwardDim { get; set; } Qwen decoder SiLU-gated feed-forward width.
int DecoderHeads { get; set; } Qwen decoder query heads.
int DecoderKeyValueHeads { get; set; } Qwen decoder key/value heads.
int EndOfTextTokenId { get; set; } Qwen’s <|endoftext|> id.
int GlobalAttentionEvery { get; set; } Every N-th ViTDet block attends globally; the rest use windows.
int ImEndTokenId { get; set; } ChatML’s <|im_end|> id, which also ends an answer.
int ImStartTokenId { get; set; } ChatML’s <|im_start|> id.
int ImageEndTokenId { get; } The </img> id.
int ImagePadTokenId { get; } <imgpad>, the image-slot token: the last id of the vocabulary.
int ImageStartTokenId { get; } The <img> id.
int NeckChannels { get; set; } Channels of ViTDet’s neck output.
int PatchSize { get; set; } ViTDet patch size, also the encoder’s stride.
int VisionMlpDim { get; set; } ViTDet MLP width.
int WindowSize { get; set; } Side of ViTDet’s windowed-attention windows.

Methods

Method Summary
void Validate() Throws if the configuration cannot build GOT-OCR2.