main-database/BELLE.yaml at preview · Language-Technology-Assessment/main-database · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
---
# Thank you for contributing!
# In filling out this yaml file, please follow the criteria as described here:
# https://osai-index.eu/contribute

# You're free to build on this work and reuse the data. It is licensed under CC-BY 4.0, with the
# stipulation that attribution should come in the form of a link to https://osai-index.eu/
# and a citation to the peer-reviewed paper in which the dataset & criteria were published:

# Liesenfeld, A. and Dingemanse, M., 2024. Rethinking open source generative AI: open-washing and the EU AI Act. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (pp. 1774-1787).

# Organization tags:
# - National origin: China
# - Contributor type: Non-academic (Company)

# Training compute:
# - Base model training compute: ~1.6e+23 FLOP +- 0.5 OoM (param count & dataset size) [Epoch AI]
# - End model training compute: unknown (likely negligible)

system:
    name: BELLE
    link: https://huggingface.co/BELLE-2/BELLE-Llama2-13B-chat-0.4M
    type: text
    performanceclass: limited
    basemodelname: Llama-2-13B
    endmodelname: BELLE-Llama2-13B-Chat-0.4M
    endmodellicense: Llama 2 Community License Agreement
    releasedate: 2023-03
    notes: Llama-2 fine-tuned on a small amount of Chinese data. Earlier variants exist based on BLOOMZ.

org:
    name: KE Technologies
    link: https://www.ke.com
    notes: Company specializing in advanced technology solutions.

# availability:
datasources_basemodel:
    class: closed
    link: https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/
    notes: Data nowhere disclosed or documented, and described only in the vaguest terms in a corporate preprint released by Meta.

datasources_endmodel:
    class: partial
    link: https://github.com/LianjiaTech/BELLE/tree/main/data/1.5M
    notes: Synthetic BELLE training data in Chinese released in batches

weights_basemodel:
    class: partial
    link: https://ai.meta.com/resources/models-and-libraries/llama-downloads/
    notes: Download only after requesting access; requires signing a consent form

weights_endmodel:
    class: partial
    link: https://github.com/LianjiaTech/BELLE/tree/main/models
    notes: Some models available, most only as delta weights requiring separate access to LLaMA

trainingcode:
    class: open
    link: https://github.com/LianjiaTech/BELLE
    notes: Repository contains a fair bit of code

# documentation
code:
    class: partial
    link: https://github.com/LianjiaTech/BELLE/blob/main/README_en.md
    notes: Quite some documentation on Github, though not all well-organized

hardware_architecture:
    class: open
    link: "https://github.com/LianjiaTech/BELLE/blob/main/README_en.md"
    notes: Specified in a fair bit of detail on GitHub.

preprint:
    class: open
    link: https://arxiv.org/abs/2303.14742
    notes:

paper:
    class: closed
    link:
    notes: No peer-reviewed paper found

modelcard:
    class: partial
    link: https://huggingface.co/BelleGroup
    notes: Provide model cards for various BELLE models on their Huggingface page; most are not fully detailed and are missing certain aspects e.g. hyperparameters.

datasheet:
    class: closed
    link:
    notes: No datasheet found.

# access:
licenses:
    class: closed
    link:
    notes: "Weights: Llama 2 Community License Agreement. Code: Apache-2.0. Data: unspecified."