notesum.ai

Published at December 5

ALMA: Alignment with Minimal Annotation

cs.CL

cs.LG

Released Date: December 5, 2024

Authors: Michihiro Yasunaga¹, Leonid Shamis¹, Chunting Zhou¹, Andrew Cohen¹, Jason Weston¹, Luke Zettlemoyer¹, Marjan Ghazvininejad¹

Aff.: ¹Meta FAIR

Arxiv: http://arxiv.org/pdf/2412.04305v1

Model

Metric: GPT-4o score

Metric: Armo score

Avg.

LIMA

Bench

Arena

Hard

Alpaca

Avg.

LIMA

Bench

Arena

Hard

Alpaca

Llama3 8B Instruct

8.25

8.78

8.06

8.37

7.48

8.59

8.41

8.74

8.07

8.59

8.06

8.63

ALMA: Our aligned models

11th round

(= final)

8.22

8.77

7.86

8.33

7.67

8.48

8.34

8.65

7.89

8.42

8.26

8.46

1st round

(= initial SFT)

7.94

8.55

7.60

8.18

7.18

8.20

7.78

8.27

7.36

7.89

7.54

7.87

Llama3 8B Base (3-shot)

4.00

4.44

3.58

4.29

3.37

4.36

3.95

4.33

3.67

4.30

3.71

3.74