Moveworks Task-Specific Models

AI model by Moveworks

High-precision, low-latency models for specialized AI tasks.

A collection of fine-tuned, high-precision machine learning models optimized for specific tasks like language detection, translation, and toxicity filtering. They provide enterprise-grade speed and reliability by handling narrow tasks that don't require large foundation models.

Model specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Languages
English

Category: Natural Language Processing

Sizes and variants
FT-LangDetection, FT-Roberta, FT-Flan-T5, FT-M2M100
Access
Integrated within Moveworks Agentic Reasoning Engine
Fine-tuning
Fine-tuned by Moveworks on reviewed usage data and feedback
Primary Architectures
RoBERTa, Flan-T5, M2M100
Language Detection Model
FT-LangDetection
Classification Model
FT-Roberta
  • Low-latency language detection with FT-LangDetection
  • Toxicity judgment for user input and assistant output via FT-Flan-T5
  • Handoff classification and entity recognition using FT-Roberta
  • Resource translation for knowledge articles via FT-M2M100

Overview

Moveworks Task-Specific Models are a collection of specialized machine learning models designed to handle narrow, high-precision tasks within the Moveworks AI platform. These models, which include fine-tuned versions of RoBERTa, Flan-T5, and M2M100, are optimized for lower latency and cost compared to large foundation models. They perform critical background operations such as language detection, toxicity filtering, entity recognition, and search relevance judgment. By using these smaller models for specific components of a request, the platform achieves higher speed and reliability for enterprise users. Moveworks maintains these models on a regular release cycle, updating them based on reviewed usage data and feedback rather than real-time conversation training.

Multilingual Capabilities

  • Language Detection: The FT-LangDetection model provides low-latency detection of the user's language to ensure the assistant responds correctly.
  • Resource Translation: The FT-M2M100 model performs two-way translation of resources such as knowledge articles and forms.

Task-Specific Operations

  • Handoff Classification: The FT-Roberta model handles handoff classification and entity recognition to identify when a request needs human intervention.
  • Search Relevance Judgment: The FT-Flan-T5 model evaluates the relevance of files found during a search to ensure helpful results.
  • Toxicity Judgment: The FT-Flan-T5 model judges the toxicity of both user input and assistant output for safe interactions.

Infrastructure and Reliability

  • Traffic Prioritization: The system prioritizes requests based on latency needs and regional or data-residency requirements.
  • Automatic Recovery: The gateway automatically recovers when a third-party endpoint is down or throttled to prevent single points of failure.

Use cases

  • Global teams use FT-LangDetection to automatically identify a user's language and respond appropriately without manual configuration.
  • Support administrators use FT-M2M100 to provide two-way translation of knowledge articles and forms for multilingual workforces.
  • Security teams rely on FT-Flan-T5 to judge the toxicity of user inputs and assistant outputs to maintain safe workplace interactions.

Specifications

Sizes and variants
FT-LangDetection, FT-Roberta, FT-Flan-T5, FT-M2M100
Access
Integrated within Moveworks Agentic Reasoning Engine
Fine-tuning
Fine-tuned by Moveworks on reviewed usage data and feedback
Primary Architectures
RoBERTa, Flan-T5, M2M100
Language Detection Model
FT-LangDetection
Classification Model
FT-Roberta
Generative Task Models
FT-Flan-T5, FT-M2M100
Update Cycle
Regular release cycle based on reviewed data
Deployment Platform
Moveworks Agentic Reasoning Engine

How it works

  1. Task Identification: The Moveworks LLM Gateway receives a request and identifies the specific task required, such as language detection or entity recognition.
  2. Model Routing: The gateway routes the request to the specialized fine-tuned model optimized for that specific task to ensure high precision and low latency.
  3. Execution and Feedback: The task-specific model processes the input and returns the result, such as a toxicity score or a translated resource, to the reasoning engine.

FAQ

Does chatting with the assistant retrain the models?

No, chatting with the assistant does not retrain the models in real time. Moveworks improves them on a regular release cycle using reviewed, masked, and anonymized usage data and feedback.

How do task-specific models differ from foundation models?

Task-specific models are smaller, faster, and easier to control than foundation models. They are trained for one specific task, such as entity recognition, providing higher quality and speed at a lower cost for that narrow use case.

How are these models evaluated?

Moveworks uses evaluation datasets curated for assistant use cases, end-to-end testing, component evaluation (like tool selection), and human annotator reviews to ensure quality before shipping.

How is model redundancy handled?

Moveworks routes traffic through an LLM Gateway that can prioritize requests by regional or data-residency requirements and automatically recovers if a specific provider is down.

Added by AntSpace from public sources (docs.moveworks.com).

Related

More AI models by Moveworks