gobean: the q4_k_m in the main model repo was hitting some tokenizer issues - after about five prompt exchanges it goes wild with repeats and unprintable tokens. I made this for comparison, it's not as bad as the q4_k_m but it isn't perfect. Both models fail the knowledge benchmark "describe the difference between a bear credit spread and a poor man's covered call," but they come really really close to getting it (just describes a standard covered call). Overall not bad. Inference is fast with q4_0 on 24gb vram - and the bug could very likely be with llama.cpp, so I may start looking into other frontends.
~ Original Model Card ~
π GitHub β’
πΎ Discord β’
π€ Twitter β’
π¬ WeChat
π Paper β’
π FAQ β’
π Learning Hub
Intro
Yi-1.5 is an upgraded version of Yi. It is continuously pre-trained on Yi with a high-quality corpus of 500B tokens and fine-tuned on 3M diverse fine-tuning samples.
Compared with Yi, Yi-1.5 delivers stronger performance in coding, math, reasoning, and instruction-following capability, while still maintaining excellent capabilities in language understanding, commonsense reasoning, and reading comprehension.
Model | Context Length | Pre-trained Tokens |
---|---|---|
Yi-1.5 | 4K | 3.6T |
Models
Chat models
Name Download Yi-1.5-34B-Chat β’ π€ Hugging Face β’ π€ ModelScope Yi-1.5-9B-Chat β’ π€ Hugging Face β’ π€ ModelScope Yi-1.5-6B-Chat β’ π€ Hugging Face β’ π€ ModelScope Base models
Name Download Yi-1.5-34B β’ π€ Hugging Face β’ π€ ModelScope Yi-1.5-9B β’ π€ Hugging Face β’ π€ ModelScope Yi-1.5-6B β’ π€ Hugging Face β’ π€ ModelScope
Benchmarks
Chat models
Yi-1.5-34B-Chat is on par with or excels beyond larger models in most benchmarks.
Yi-1.5-9B-Chat is the top performer among similarly sized open-source models.
Base models
Yi-1.5-34B is on par with or excels beyond larger models in some benchmarks.
Yi-1.5-9B is the top performer among similarly sized open-source models.
Quick Start
For getting up and running with Yi-1.5 models quickly, see README.