AgriBERT Triton Configuration

This repository contains the NVIDIA Triton Inference Server configuration for serving the AgriBERT model (recobo/agriculture-bert-uncased) using the PyTorch LibTorch backend. The repository currently focuses on the model configuration (config.pbtxt), which defines inputs, outputs, and execution settings required to deploy the model on Triton.

📌 Model Overview

undefinedAgriBERT is a BERT-based language model trained specifically for the agriculture domain.

• Base model: SciBERTundefined
• Further pre-trained using Masked Language Modeling (MLM)undefined
• Domain: Agriculture research and practical agricultural knowledge
• Language: English

Training Data

The model was trained on a large agriculture-focused corpus consisting of:

• undefined1.2 million paragraphs from the US National Agricultural Library (NAL)
• undefined5.3 million paragraphs from agriculture-related books and common literature

This balanced dataset includes both scientific and general agricultural content, enabling strong domain-specific language understanding.

Training Objective

• undefinedMasked Language Modeling (MLM)undefined
• 15% of input tokens are randomly masked
• The model predicts the masked tokens using bidirectional context

This allows AgriBERT to learn deep contextual representations relevant to agricultural text.

🧠 Use Case

While AgriBERT is commonly used for fill-mask tasks, this Triton configuration exposes the model as a general-purpose inference endpoint, suitable for:

• Text classification
• Feature extraction
• Downstream agriculture NLP tasks

The output shape indicates a 7-class FP32 output, suggesting use in a classification or scoring setup.

⚙️ Triton Model Configuration

The config.pbtxt defines how Triton serves the model.

Model Details

• undefinedModel Name: agribert
• undefinedBackend: pytorch_libtorch
• undefinedExecution: GPU-based inference

Inputs

Name	Data Type	Shape	Description
input_ids	INT32	[batch_size, seq_len]	Token IDs from the tokenizer
token_type_ids	INT32	[batch_size, seq_len]	Segment IDs for sentence pairs
attention_mask	INT32	[batch_size, seq_len]	Attention mask for padding tokens

All inputs support dynamic batch size and dynamic sequence length.

Outputs

Name	Data Type	Shape	Description
OUTPUT__0	FP32	[batch_size, 7]	Model logits or class scores

Instance Group

• undefinedGPU Instances: 1
• undefinedKind: KIND_GPU

📂 Repository Structure

.
├── agribert/
│   ├── config.pbtxt
│   └── 1/
│       └── model.pt  (expected)
└── README.md

⚠️ Note: The actual serialized TorchScript model (model.pt) must be placed under a versioned directory (e.g., 1/) for Triton to load the model successfully.

🚀 Deployment Notes

• Ensure the model is exported to TorchScript (torch.jit.trace or torch.jit.script)
• Tokenization must be handled client-side using the same tokenizer:
  • recobo/agriculture-bert-uncased
• Input tensors must be provided as INT32undefined

🔗 Model Reference

• Hugging Face Model: recobo/agriculture-bert-uncasedundefined
• Task: Masked Language Modeling (MLM)
• Framework: PyTorch / Transformers

undefined📄 License & Attribution: All credit for the model architecture and pretraining goes to Recobo.ai and the original AgriBERT authors. This repository only contains deployment configuration for NVIDIA Triton Inference Server.