Skip to content

Latest commit

 

History

History
84 lines (58 loc) · 2.69 KB

openhermes.md

File metadata and controls

84 lines (58 loc) · 2.69 KB

Overview

OpenHermes 2.5 Mistral 7B is a state-of-the-art Mistral fine-tune, representing a continuation of the OpenHermes 2 model with additional training on code datasets. It demonstrates exceptional instruction-following capabilities.

Training Data

The model was trained on 1,000,000 entries comprising:

  • Primarily GPT-4 generated data
  • High-quality data from open datasets across the AI landscape
  • Code instruction datasets (7-14% of total)
  • Diverse task types and domains

Key Innovation

Training on a good ratio of around 7-14% of the total dataset of code instruction boosted several non-code benchmarks, including:

  • TruthfulQA: Improved truthfulness
  • AGIEval: Better reasoning
  • GPT4All suite: Enhanced general capabilities

Performance

OpenHermes leads in summary generation tasks, particularly excelling at:

  • Consistency in summaries
  • Minimizing contradictions
  • Instruction-following accuracy
  • Code generation and explanation

Key Features

  • High-Quality Training: 1M GPT-4 generated examples
  • Code-Enhanced: Better overall performance through code training
  • Instruction Excellence: Strong instruction-following
  • Diverse Capabilities: General-purpose applications
  • Compact Size: 7B parameters

Training Methodology

  • Fine-tuned from Mistral-7B base
  • Curated high-quality instruction dataset
  • Balanced mix of code and non-code tasks
  • GPT-4 generated examples for quality

Supported Tasks

  • Instruction Following: Precise task execution
  • Code Generation: Programming assistance
  • Text Summarization: Consistent, contradiction-free summaries
  • Question Answering: Accurate information retrieval
  • Conversational AI: Natural dialogue
  • Content Generation: Creative and technical writing

Performance Highlights

Summarization: OpenHermes and Starling are in the lead, depending on whether higher importance is given to consistency or minimizing contradictions.

General Benchmarks: Improved scores across multiple evaluation suites due to code training integration.

Use Cases

  • Development assistance with code generation
  • Document summarization
  • Instruction-based task automation
  • Conversational interfaces
  • Content creation and editing
  • Educational applications

Deployment

  • Efficient inference on consumer hardware
  • Quantization support (AWQ and other methods)
  • Compatible with vLLM for fast serving
  • Standard Hugging Face integration

Comparison

OpenHermes competes strongly with other instruction-tuned 7B models, often leading in specific tasks like summarization while maintaining broad capabilities.

Licensing

Based on Mistral-7B, follows Apache 2.0 license terms.

Pricing

Free and open-source.