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.
FWIW, Claude Opus (the paid model) gets the correct answer, and explains it well:
Based on the information provided, we cannot definitively determine whether cherries are better than bananas. The statements given only provide relative comparisons between apples and bananas, and apples and cherries, but do not directly compare cherries and bananas.
Here's what we know:
1. Apples are better than bananas.
2. Cherries are worse than apples.
However, these statements do not necessarily imply a transitive relation. In other words, just because A is better than B, and C is worse than A, it doesn't always mean that C is worse than B.
For example, if we assign numerical values to represent the quality of each fruit (higher numbers indicate better quality), we could have a scenario like this:
- Apples: 8
- Bananas: 6
- Cherries: 7
In this case, apples (8) are better than bananas (6), and cherries (7) are worse than apples (8), but cherries (7) are still better than bananas (6).
Therefore, more information would be needed to determine the relative quality of cherries compared to bananas.
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.
This makes sense to me. If you think about the training data, texts working through problems using formal predicate logic are likely to be correct, and much more likely to be precise about what information is (or isn’t) contained in the propositions. So if you formulate the problem in this language, you’re prompting the model to sample from patterns that are more likely to give you the result you want. Whereas if you use regular English, it could be sampling from cooking blogs or who knows what.
- 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/zephyr-orpo-141b-A35b-v...
- Base Model: https://huggingface.co/mistral-community/Mixtral-8x22B-v0.1
- Fine-tune data: https://huggingface.co/datasets/argilla/distilabel-capybara-...
- Recipe/code to train the model: https://huggingface.co/datasets/argilla/distilabel-capybara-...
- Open-source inference engine: https://github.com/huggingface/text-generation-inference
- Open-source UI code https://github.com/huggingface/chat-ui
Have fun!