NEW! AutoTune™ From Armilla AI
AutoTune™ enables robust Generative AI solutions with automated alignment
Harden your Generative AI solutions with advanced testing and fine-tuning
to increase performance and reduce bias, hallucinations and safety issues.
Benefits of AutoTune
-
Immediate Value
A wide range of out-of-the-box alignment controls available, with the ability to customize or create new controls. -
High Performance
Advanced fine-tuning, enabling greater robustness than traditional approaches. -
No Data Required
Concept-based approach requires minimal data, and minimal human-in-the-loop contact. Auto generates targeted synthetic data. -
Automated Testing
Each alignment scenario can be tested and assessed independently. Detect bias, security holes, tonality issues, hallucinations easily. -
Supports Major Base Models
Including OpenAI LLMs like GPT, Stable Diffusion image generation models, open source models, with Bard and others coming soon.
One Framework To Rule Them All
Test, compare, tune and select from a wide range of models
allowing for flexible deployment, cost and performance options
STEP 1
Define:
Use out-of-the-box alignment controls or define your own performance expectations using concepts (not datasets), enabling very broad alignment to your goals.
Example use cases:
STEP 2
Synthesis:
AutoTune™ takes it from there, synthesizing data and testing to determine model robustness. Permutations can then be used for robustness testing and fine-tuning models.
STEP 3
Interpret:
The 'fingerprinting' technique drills into model performance to understand and prepare to optimize for aligned model behavior.
STEP 4
Fine-tune:
AutoTune™ enhances models using a new technique to optimize fine-tuning based on concepts, not data.
STEP 5
Feedback:
AutoTune™ can automate the feedback loop and repeat the process, or loop back in a human to assess performance.
Currently Supported Alignment Controls
| Class | Functionality | Description | Control Type |
|---|---|---|---|
| Security | Privacy filtering | Redacts PII and sensitive information, preventing leakage to / from the base model | Guardrail |
| Fact checking | Provides a framework for detecting hallucinations and validating information returned from the model | Guardrail | |
| Jailbreaks | Provides robustness testing, fine-tuning and prevention of common hacking attempts | Hybrid | |
| Toxicity | Provides detection and interception of toxic inputs & outputs | Guardrail | |
| Bias | Stereotypical tuning | Change stereotypical conventions and mitigate biases from training data that has been incorporated into the base model | Fine-tuning |
| Ethical filtering rules | Provide and fine-tune with additional context to ensure context-appropiate responses | Hybrid | |
| Custom | Custom datasets | Allows custom datasets to be used for answering and guardrails around its answers | Hybrid |
| Transfer learning and custom classification | Allows development of custom training of default classifications as well as smaller / more secure models | Fine-tuning | |
| Prompt optimization | Optimizes responses via generated prompt engineering | Guardrail | |
| Tonality | Provides fine-tuning for adjusting the tonality and behavior of the model | Fine-tuning |
