Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

osanseviero wrote: post #2

Zephyr 141B is a Mixtral 8x22B fine-tune. Here are some interesting details - Base model: Mixtral 8x22B, 8 experts, 141B total params, 35B activated params - Fine-tuned with ORPO, a new alignment algorithm with no SFT step (hence much faster than DPO/PPO) - Trained with 7K open data instances -> high-quality, synthetic, multi-turn - Apache 2 Everything is open: - Final Model: https://huggingface.co/HuggingFaceH4/zeph…

I like that they say how the model was trained for 1.3 hours on 4 nodes of 8 x H100s. By my rough calculation, that should probably have cost around $100 or so. (At $2 per hour, x 8 gpus x 4 nodes). Not free, but pretty cheap in the scheme of things. At least, once you know what you're doing.

Re: Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

#4

Post by dloss » Fri, Apr 12, 2024, 5:32 PM UTC

osanseviero wrote: post #2

Zephyr 141B is a Mixtral 8x22B fine-tune. Here are some interesting details - Base model: Mixtral 8x22B, 8 experts, 141B total params, 35B activated params - Fine-tuned with ORPO, a new alignment algorithm with no SFT step (hence much faster than DPO/PPO) - Trained with 7K open data instances -> high-quality, synthetic, multi-turn - Apache 2 Everything is open: - Final Model: https://huggingface.co/HuggingFaceH4/zeph…

I wanted to write that TGI inference engine is not Open Source anymore, but they have reverted the license back to Apache 2.0 for the new version TGI v2.0: https://github.com/huggingface/text-generation-inference/rel...

Good news!

Re: Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

#5

Post by leblancfg » Fri, Apr 12, 2024, 5:54 PM UTC

osanseviero wrote: post #2

Zephyr 141B is a Mixtral 8x22B fine-tune. Here are some interesting details - Base model: Mixtral 8x22B, 8 experts, 141B total params, 35B activated params - Fine-tuned with ORPO, a new alignment algorithm with no SFT step (hence much faster than DPO/PPO) - Trained with 7K open data instances -> high-quality, synthetic, multi-turn - Apache 2 Everything is open: - Final Model: https://huggingface.co/HuggingFaceH4/zeph…

What does ORPO stand for? Can't seem to find related links.

Re: Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

#6

Post by cateye » Fri, Apr 12, 2024, 6:04 PM UTC

leblancfg wrote: post #5

osanseviero wrote: post #2

Zephyr 141B is a Mixtral 8x22B fine-tune. Here are some interesting details - Base model: Mixtral 8x22B, 8 experts, 141B total params, 35B activated params - Fine-tuned with ORPO, a new alignment algorithm with no SFT step (hence much faster than DPO/PPO) - Trained with 7K open data instances -> high-quality, synthetic, multi-turn - Apache 2 Everything is open: - Final Model: https://huggingface.co/HuggingFaceH4/zeph…

What does ORPO stand for? Can't seem to find related links.

Odds Ratio Preference Optimization (ORPO): https://arxiv.org/abs/2403.07691

Re: Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

#7

Post by mjewkes » Fri, Apr 12, 2024, 6:52 PM UTC

My current favorite “LLM breaker” below. GPT4, Claude, and this all fail.

—-

Apples are better than bananas. Cherries are worse than apples. Are cherries better than bananas?

Re: Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

#8

Post by loudmax » Fri, Apr 12, 2024, 7:33 PM UTC

mjewkes wrote: post #7

My current favorite “LLM breaker” below. GPT4, Claude, and this all fail. —- Apples are better than bananas. Cherries are worse than apples. Are cherries better than bananas?

To be fair, a lot of humans fail that. Including people that should know better.

Re: Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

#9

Post by mjewkes » Fri, Apr 12, 2024, 8:51 PM UTC

loudmax wrote: post #8

mjewkes wrote: post #7

My current favorite “LLM breaker” below. GPT4, Claude, and this all fail. —- Apples are better than bananas. Cherries are worse than apples. Are cherries better than bananas?

To be fair, a lot of humans fail that. Including people that should know better.

For sure. It's not a fair prompt at all. I'm super bullish on LLMs and am using GPT-4 in production right now. This stuff is magic.

It's actually hard to find short, simple, "plain english" failure cases like the above.

The "chain of reasoning" that the modern models deploy before the fail is funny too. This is GPT-4:

---

To determine the relationship between cherries and bananas based on your statements, let's break it down:

  1. Apples are better than bananas.
  2. Cherries are worse than apples.

From statement 1, we know apples rank higher than bananas. Statement 2 tells us cherries rank lower than apples. By this logic, since cherries are lower than apples, which are higher than bananas, it follows that cherries are also lower than bananas.

Therefore, based on these comparisons, cherries are not better than bananas.

Re: Zephyr 141B, a Mixtral 8x22B fine-tune, is now available in Hugging Chat

#10

Post by mjewkes » Fri, Apr 12, 2024, 8:56 PM UTC

mjewkes wrote: post #9

loudmax wrote: post #8

Earlier quoted context omitted.

To be fair, a lot of humans fail that. Including people that should know better.

For sure. It's not a fair prompt at all. I'm super bullish on LLMs and am using GPT-4 in production right now. This stuff is magic. It's actually hard to find short, simple, "plain english" failure cases like the above. The "chain of reasoning" that the modern models deploy before the fail is funny too. This is GPT-4: --- To determine the relationship between cherries and bananas based on your statements, let's break…

Notably, if you ask it to transform the statements to formal logic, you get a correct response! This stuff is truly magic.

https://chat.openai.com/share/81e45fef-a72b-4258-98d6-5c8190...