We recently released Smaug-72B-v0.1 which has taken first place on the Open LLM Leaderboard by HuggingFace. It is the first open-source model to surpass an average score of 80%.
We created Smaug-72B-v0.1 using a new fine-tuning technique, DPO-Positive (DPOP), and new pairwise preference versions of ARC, HellaSwag, and MetaMath (as well as other existing datasets). We introduce the technique and the full training details in our new paper: https://arxiv.org/abs/2402.13228.
We show that on datasets in which the edit distance between pairs of completions is low (such as in math-based datasets), standard DPO loss can lead to a reduction of the model's
likelihood of the preferred examples, as long as the relative probability between the preferred and dispreferred classes increases.
Using these insights, we design DPOP, a new loss function and training procedure which avoids this failure mode.
Surprisingly, we also find that DPOP outperforms DPO across a wide variety of datasets and downstream tasks, including datasets with high edit distances between completions.
We believe this new approach is generally useful in training across a wide range of model types and downstream use cases, and it powers all of our Smaug models.
With the release of our paper and datasets, we are excited for the open source community to continue to build on and improve Smaug and spawn more dragons to dominate the LLM space!
Evaluation Results
Average
ARC
HellaSwag
MMLU
TruthfulQA
Winogrande
GSM8K
80.48
76.02
89.27
77.15
76.67
85.08
78.70
MT-Bench
We ran MT-Bench with the llama-2 conversation template and the system prompt set to the Qwen system prompt. We got the following
results in single model mode:
First Turn
Second Turn
Average
8.18
7.34
7.76
The bottom of the card includes some sample responses from the model. We have selected responses that have high and low scores
from the GPT-4 judge.
By comparison, MoMo-72B-lora-1.8.7-DPO has the following results:
ARC
TruthfulQA
GSM8K
0.20
0.39
1.00
Note that GSM8K often scores very highly on this contamination suite - we verified this by also running Llama-2-70B:
ARC
TruthfulQA
GSM8K
0.22
0.51
0.89
Sample MT Bench results
Example 1
Question
Imagine you are writing a blog post comparing two popular smartphone models.
Develop an outline for the blog post, including key points and subheadings to effectively
compare and contrast the features, performance, and user experience of the two models.
Please answer in fewer than 200 words.
Response
Outline:
Introduction to the two smartphone models being compared (mention brand names and models)
Design and Build Quality
Comparison of dimensions, weight, materials used, and overall aesthetic appeal
Water and dust resistance, if applicable
Display and Screen Quality
Comparison of screen technology (OLED, IPS LCD, etc.)
Display resolution, refresh rate, and brightness levels
Performance and Hardware
Processor, RAM, and storage configurations
Comparison of benchmark scores (e.g., AnTuTu, Geekbench)
Camera System and Image Quality
Comparison of primary, ultra-wide, telephoto, and macro lenses (if available)
Comparison of low-light performance and image processing capabilities
Battery Life and Charging
Battery capacity and estimated usage time
Charging technologies (wired, wireless, fast charging, reverse charging)
Software and User Experience
Comparison of operating systems (iOS vs. Android) and their features
User interface, customization options, and app ecosystem
Additional Features and Extras
Comparison of biometric authentication (fingerprint, face unlock, etc.)
Additional features like NFC, 5G connectivity, and audio quality
Pricing and Value for Money
Comparison of retail prices and current offers
Assessment of overall value considering the features and performance
Conclusion and recommendation for potential buyers based on their preferences
Example 2
Question
In a survey conducted at a local high school, preferences for a new school color were measured:
58% of students liked the color blue, 45% preferred green, and 22% liked both colors.
If we randomly pick a student from the school, what's the probability that they would like neither blue nor green?
Answer
To find the probability that a randomly picked student would like neither blue nor green, we need to find the percentage of students who like only one color or neither color.
From the given information, we know:
58% like blue
45% prefer green
22% like both colors
To find the percentage of students who like only one color, we can subtract the percentage of students who like both colors from the sum of students who like blue and green:
58% (blue) + 45% (green) - 22% (both) = 83%
Now, to find the probability of a student liking neither blue nor green, we subtract the percentage of students who like at least one color from 100%:
100% - 83% = 17%
So, the probability that a randomly picked student would like neither blue nor green is 17%.
Example 3
Question
Implement a program to find the common elements in two arrays without using any extra data structures.
Answer
Finding common elements in two arrays without using extra data structures can be achieved by using a simple algorithm. Here's a Python implementation of this problem:
python
1deffind_common_elements(arr1, arr2):2 common_elements =[]3for num1 in arr1:4if num1 in arr2 and num1 notin common_elements:5 common_elements.append(num1)6return common_elements
78# Example usage:9array1 =[1,2,3,4,5]10array2 =[4,5,6,7,8]11print(find_common_elements(array1, array2))# Output: [4, 5]
This program iterates through the elements of the first array (arr1) and checks if each element is present in the second array (arr2). If it is, and the element hasn't been added to the common_elements list yet, it is appended. The function then returns the list of common elements.
Example 4
Question
Identify the named entities (people, organizations, locations) mentioned in the given news article.
Please generate a JSON dictionary that lists the named entities in three separate groups based on their entity types.
The key is the type of entity and the value is a list of strings.
Yesterday, Adamson Emerson, the CEO of Faraday, and Dieter Zetsche, the CEO of Daimler AG, announced plans to
build a new Gigafactory in Berlin. The facility will be a joint venture between Faraday and Daimler, producing
electric vehicles and battery packs for both companies, creating thousands of job opportunities in the region.
Emerson and Zetsche stated that the strategic location of Berlin, coupled with its skilled workforce and strong infrastructure,
makes it an ideal choice for expansion. The new Gigafactory aims to meet the growing demand for electric vehicles in Europe
and contribute to a sustainable future. Volkswagen CEO Herbert Diess welcomed the news, saying greater collaboration will,
benefit the auto industry's transition to e-mobility.
Please cite the paper if you use data, model, or method in this repo.
@article{pal2024smaug,
title={Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive},
author={Pal, Arka and Karkhanis, Deep and Dooley, Samuel and Roberts, Manley and Naidu, Siddartha and White, Colin},
journal={arXiv preprint arXiv:2402.13228},
year={2024}
}
