Yi 1.5
github.com
github.com
They have been releasing a lot of really good models over the last ~6 months. Their previous (1.0?) Yi-34B-Chat model ranks similar to GPT-3.5 on Chatbot Arena. [1] A quantized version of that model can be run on a single consumer video card like the RTX 4090.
This new set of models should raise the bar again by adding more options to the open source LLM ecosystem. If you inspect the config.json[2] in the model repo on HuggingFace, you can see that the model architecture is LlamaForCausalLM (the same as Meta's Lllama). The difference between the Yi models and a simple fine-tuning is that Yi models have had a different set of data, configuration, and process going back to the pre-training stage.
Their models perform well in Chinese and in English.
There are a lot of good models coming out of China, some of which are only published to ModelScope. I haven't spent much time on ModelScope because I don't have a Chinese mobile number to use to create an account. Fortunately, Yi publish to HuggingFace as well.
[1] https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboar...
[2] https://huggingface.co/01-ai/Yi-1.5-34B-Chat/blob/fa695ee438...
What other values and biases have been RLHFed there and for what purpose?
What happened at Tienanmen square?
yi-34b-chat:
Tiananmen Square is a large public square in the heart of Beijing, China. It is named after the Tiananmen (literally "Gate of Heavenly Peace"), which is the gate located to the north of the square and was the main entrance to the Imperial City during the Ming and Qing dynasties.
The square is famously known for two significant events in Chinese history:
The May Fourth Movement (1919):
This was a student-led movement that began on May 4, 1919, in response to the Chinese government's handling of the Treaty of Versailles, which granted German concessions in Shandong to Japan rather than returning them to China. The movement sparked a wave of anti-imperialist sentiment and calls for cultural and political reform, and it is considered a key event leading to the birth of Chinese nationalism and the development of the Chinese Communist Party.
The Tiananmen Square Massacre (1989):
This is the event for which Tiananmen Square is most internationally renowned. In the spring of 1989, there were student-led demonstrations in Beijing and other cities across China, calling for political reform and an end to corruption. The protests were initially peaceful, but on June 3 and 4, the Chinese government responded with a military crackdown, opening fire on the protesters and bystanders, killing and injuring thousands. The exact death toll is unknown, but it is estimated to be in the hundreds or even thousands.
I’m going to guess a Chinese model would have a different response, and GPT-4 has been “aligned” to lie about uncomfortable facts.
I wouldn't put that on the same level as "refusing to talk about massacre of civilians"; but I wouldn't put it to the level of "free and unbiased" either.
Asking if Palestinians deserve freedom was answered with something like " It's a complex and sensitive topic".
https://www.reddit.com/r/Palestine/comments/181gi73/chatgpt_...
I tested it again just now, same result.
And in any case, whatever model you train is going to have the biases of the training datasets, and if you make heavy use of Wikipedia you will have the footprint of Wikipedia in your output, for good or bad.
Where can I find LLMs that are uncensored and accessible (without logging in)?
I remember, because I spent non-trivial effort trying to make it work for long-form technical summarization. My lackluster findings were validated by RULER.
This will ban Chinese characters from the sampling process. Works for Yi and Qwen models.
1: https://arxiv.org/html/2403.04652v1
2: https://github.com/openai/gpt-3/blob/master/dataset_statisti...
The small GLM models were like 50-50 English-Chinese in pretraining but much more Chinese in instruct training. Had the same issue until they balanced that.
Yi 34b with results similar to Llama 3 70b and Mixtral 8x22b
Yi 6b and 9b with results similar to Llama 3 8b
The "will it tell me how to make meth" stuff is a huge source of noise, which you could argue is digging for refusals which can be annoying, and the benchmark claims to filter out... but in reality a bunch of the refusals are soft refusals that don't get caught, and people end up downvoting the model that's deemed "corporate".
Honestly the fact that any closed source model with guardrails can even place is a miracle, in a proper benchmark the honest to goodness gap between most closed source models and open source models would be so large it'd break most graphs.
There aren't nearly as many 3.5 level models as the leaderboard implies for example.
LLM benchmarks are horribly broken. IMHO there is better signal in just looking at parameter counts.
I'm considering a new laptop later this year and the ram is now fixed to 16GB on most of them.
I plan on digging deep into ML during my coming break from paid work.
I'll keep this in mind!
I recently bought a 7900 XTX with 24 GB of VRAM, but the model I currently run can easily run in 16 GB (6 bit llama 3 8b). It's fast enough and high enough quality that I can use it for processing information that I don't feel comfortable sharing with hosted services. It's definitely not the best of the best as far as what models are able to do right now, but it's surprisingly useful.
Unless of course you were talking about VRAM, in which case 16GB is still not great for ML (to be fair, the 24GB of an RTX 4090 aren't either, but there's not much more you can do in the space of consumer hardware). I don't think the other commenter was talking about VRAM, because 16GB VRAM are very overkill for everyday computing... and pretty decent for most gaming.
You don’t need a GPU for llm inference. Might not be as fast as it could be but usable.
I don't want to have a laptop over 3 pounds and I'm not spending over 1100$, so a dedicated GPU isn't really an option.
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.
Yi-1.5 comes in 3 model sizes: 34B, 9B, and 6B. For model details and benchmarks, see Model Card."
Literally after that...