MythoOrca2_exl2_4bpw

MythoOrca Model Card

Overview

Model Name: MythoOrca
Type: Language Model
Version: 4-bit Quantization
Mixture: 50% MythoMax and 50% Orca 2
Developer: Combination of two distinct AI research initiatives

Intended Use

• Research: MythoOrca is primarily intended for research purposes, combining the reasoning abilities of Orca 2 with the storytelling and roleplaying capabilities of MythoMax.
• Roleplaying and Storywriting: Given its heritage from MythoMax, it excels in these areas.
• Reasoning and Comprehension: Inherits Orca 2's strengths in reasoning, reading comprehension, and problem-solving.

Model Architecture

• Base Models: Combines LLAMA-2 (Orca 2 base) and Huginn (from MythoMax).
• Layer Composition: Features a unique tensor type merge technique, with each of its 363 tensors having a unique ratio applied to it.

Quantization

• Bit-Depth: 4-bit, optimizing the model for more efficient computation while maintaining performance.
• Source: Quantized models facilitated by TheBloke's GPTQ (GGML - GPTQ).

Performance

• Capabilities: Enhanced coherence across the model structure, proficient in roleplaying, storywriting, and reasoning tasks.
• Limitations: As a quantized model, some nuances in performance compared to its full-bit counterparts might be observed.

Data and Training

• Training Data: Synthetic dataset (Orca 2) and a blend of MythoLogic-L2 and Huginn data (MythoMax).
• Training Method: A combination of fine-tuning and an experimental tensor merge technique.

Bias, Risks, and Limitations

• Data Biases: Inherits biases from both parent models, potentially reflecting biases in source data.
• Contextual Understanding: Limited real-world understanding, potentially leading to inaccuracies.
• Transparency: Complex architecture makes it difficult to understand specific output rationales.
• Content Harms: Potential to generate biased or harmful content. Usage of content moderation services recommended.
• Hallucination: Risk of fabricating content, especially in complex or ungrounded scenarios.
• Misuse Potential: Risk of being used to generate disinformation or harmful content.

Usage Guidelines

• Intended for Research: Not optimized for downstream applications without further analysis.
• Content Moderation: Integration with content safety tools like Azure AI Content Safety recommended.
• Prompt Format: Best used with Alpaca formatting for optimal performance.

Licensing

• Orca 2 License: Microsoft Research License.
• LLaMA 2 License: LLAMA 2 Community License, Copyright © Meta Platforms, Inc.

Additional Information

• Model Availability: Publicly accessible for research purposes.
• Documentation and Support: Refer to the original papers and technical reports of MythoMax and Orca 2 for detailed information.

Getting Started Example

Refer to Orca 2's implementation with Hugging Face's transformers library for a basic usage example, adapted for MythoOrca's unique capabilities.

---

license: llama2