redactable-dolphin-mixtral

Redactable-LLM

The high-level overview for integrating multiple Open Source Large Language Models within the AutoGen Framework is as follows:

Development of Custom Agents

• Agent Design: Tasks include NLP/NER/PII identification, interpreting natural language commands, executing document redaction, and final verification.
• Customization: Custom agents trained on specific tasks related to each aspect of the redaction process.
• Human Interaction: Implement features to facilitate seamless human-agent interaction, allowing users to input commands and queries naturally (Optional)

LLM & VLLM AutoGen Integration

• Model Selection: Automatic, task-dependent agent selection.
• Enhanced Inference: Enhanced LLM inference features for optimal performance, including tuning, caching, error handling, and templating.
• Quality Control: Vision agents analyze redacted documents using Set-of-Mark (SoM) prompting. Rejected documents are reprocessed and reviewed.

System Optimization

• Workflow Automation: Automate the redaction workflow using a blend of LLMs, custom agents, and human inputs for efficient detection and redaction of sensitive information.
• Performance Maximization: Optimize the system for both efficiency and accuracy, utilizing AutoGen's complex workflow management features.

User Interface Development

• Interface Design: Develop a user-friendly interface that enables non-technical users to interact with the system via natural language prompts.
• Feedback Integration: Implement a feedback loop to continuously refine the system's accuracy and user-friendliness based on user inputs.
• User Knowledgebase: (Optional) User account, profile, and domain knowledge will be accessible by the Research agent, for personalized interaction and results.

Training, Testing and Validation

• Model Training: Develop new datasets, focused on document understanding related to redaction.
• Unit Testing: Conduct extensive unit tests to ensure individual system components function correctly.
• System Testing: Perform comprehensive end-to-end testing to validate the entire redaction process, from user input to output.
• User Trials: Facilitate user trials to gather feedback and make necessary system adjustments.

---

• Mistral AI (LLM)

  Paper | Model

• QwenLM (VLLM)

  Paper | Code | Paper: Set-of-Mark Prompting

• AutoGen

  Paper | Code

• Gretel AI (Synthetic Dataset Generation)

  Model Page | Code | Paper: Textbooks Are All You Need II