Yi-1.5 is a suite of large language models (LLMs) offering enhanced capabilities in coding, math, reasoning, and instruction following. Targeting developers and researchers, it provides improved performance over its predecessor, Yi, with models available in 34B, 9B, and 6B parameter sizes.

How It Works

Yi-1.5 is built upon a foundation of continuous pre-training on a 500 billion token corpus, followed by fine-tuning on 3 million diverse samples. This extensive training regimen aims to bolster its proficiency in complex cognitive tasks while retaining strong language understanding and commonsense reasoning abilities.

Quick Start & Requirements

• Installation: pip install -r requirements.txt
• Prerequisites: Python 3.10+, transformers library. Models can be downloaded from Hugging Face, ModelScope, or WiseModel.
• Local Inference: Example provided using Hugging Face transformers for local execution on CUDA-enabled GPUs.
• Ollama: Supports running Yi-1.5 models locally via ollama run yi:v1.5.
• vLLM: Deployment via vLLM's OpenAI-compatible API server is supported.
• Web Demo: Local web demo available via python demo/web_demo.py -c <your-model-path>.
• Docs: Yi Cookbook, FAQ, Learning Hub

Highlighted Details

• Enhanced performance in coding, math, reasoning, and instruction-following.
• Available in 34B, 9B, and 6B parameter sizes.
• OpenAI-compatible API available via Yi Platform, Replicate, and OpenRouter.
• Supports fine-tuning with popular frameworks like LLaMA-Factory, Swift, XTuner, and Firefly.

Maintenance & Community

• Active community support via Discord and Twitter.
• Discord, Twitter

Licensing & Compatibility

• Licensed under Apache 2.0.
• Derivative works require attribution to 01.AI.

Limitations & Caveats

The README does not specify hardware requirements for each model size or provide explicit benchmarks comparing Yi-1.5 against other leading LLMs.