I compared different open source models and discovered that Yi-34B-chat model was exceptional in its answers. I spent more time chatting with it and compare it with mixtral 8x7b, cloud 2, gpt 3.5 turbo, and llama 2. I found that yi-34b generated more diverse and creative responses than the others. The only close os competitor was mixtral 8x7b, however mixtral had more hallucinations and tend to use more complex words than its counterpart. I suggest using yi-34b chat model on myshell site, as it matches the site's focus on role play and storytelling scenarios. The model is very diverce and unique in its responses and its word choices. I tested the model with different kinds of creative content, such as poems, lyrics, stories, dialogues, nursery rhymes, and plays. The model was very responsive to the user’s choice of tone and showed different ranges of literary tones and styles. It also maintained consistency in the story plot, characterization, and had creativity in introducing plot twists and surprising events. One thing that impressed me was that yi-34b could produce full stylistic creative texts with minimal instructions by the user. The model showed great performance in all these queries. All the yi models are unbiased and uncensored, suitable for creative tasks. I did not test the model's mathematical or coding abilities, but the model has shown great performance in all the benchmarks as evaluated in llm testing platforms. In this thread, I will briefly introduce yi-34b, provide some links and some news mentioning some controversies. You can chat test the chat model here or in replicate with more adjustable configurations.
Contact info:
01.AI: linkedin Github , huggingface. And X.
Dr Kai-Fu lee : linkedin , twitter
The yi-34b chat model, which was released to the public on november 23th with an initial 4k token memory, is a fine-tuned version of the Chinese-English bilingual model yi-34b. The yi-34b base model, which was published in november 3th, is based on the LLaMA architecture and developed by 01.AI. 01.AI recently developed some quantized models that can run on home desktops with at least 20g of VRAM. There are different fine-tune versions based on yi, such as nous hermes 2 yi-34b, sus-chat and dolphin-2.2-yi-34b-200k . However, there are some fine tune models with biased dats and There was some fine-tune versions that was trained on uncensored and unbiased data selectively chosen by the developer.
Yi-34B chat model is a fine-tuned version of the Chinese-English bilingual model yi-34b. It was published on November 3th and had since great scores in all the benchmarks. It was ranked 2nd in the Alpaca Eval (currently 6th). and 9th in the SuperClue Chinese language benchmark. It is 6th in the Helm Leaderboard and 7th in the LMSys Org Leaderboard, (before Tulu). It has 1110 elo rating in the Chatbot Arena Leaderboard on HuggingFace. Yi-34B chat model outperformed GPT 3.5 turbo in almost all the leaderboards. Nous Hermes 2 yi-34b, which is a fine-tuned version based on yi, became the best of the Hermes series. 01.AI is constantly improving their models. These results show a capable base model and an accurate training process, and a promising future for open source models and community. I personally found yi-34b creative, less hallucinatory than mixtralx87b, and more coherent. It is also free.
On November 6th, Eric Hartford developer of Dolphine models, raised this issue that yi-34b architecture is identical to LLaMA with only two tensors renamed. This led to a lot of criticism against 01.AI in the Chinese media and the open source community. 01.AI Developer Richard Lin responded to this issue and said that it was an oversight during some experimental training. However, the Yi series models are NOT derivatives of LLaMA, as they do not use LLaMA's weights. .The Yi model is an outstanding open-source model, containing a large amount of original training techniques and datasets, which are the intellectual property of Yi company. So yi-34B remains under 01.AI licence. Read this article in medium. Kai-Fu Lee's 01.AI Startup Addresses Controversy Surrounding Model. And **Yi relation to llama.**
“ The Yi series models are fully open for academic research and free commercial usage with permission via applications. All usage must adhere to the Yi Series Models Community License Agreement 2.1. For free commercial use, you only need to send an email to get official commercial permission. “