TianyiQ commited on
Commit
0e3df26
1 Parent(s): 1127533

Upload ./README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +72 -28
README.md CHANGED
@@ -1,39 +1,57 @@
1
  ---
2
- license: other
3
- base_model: meta-llama/Meta-Llama-3-70B
4
- tags:
5
- - llama-factory
6
- - full
7
- - generated_from_trainer
8
- model-index:
9
- - name: C019_random_sample_Meta-Llama-3-70B_instruct_20240512_233548
10
- results: []
11
  ---
12
 
13
- <!-- This model card has been generated automatically according to the information the Trainer had access to. You
14
- should probably proofread and complete it, then remove this comment. -->
15
 
16
- # C019_random_sample_Meta-Llama-3-70B_instruct_20240512_233548
17
 
18
- This model is a fine-tuned version of [./output/training_results/C019_random_sample_Meta-Llama-3-70B_pretrain_20240512_020339](https://huggingface.co/./output/training_results/C019_random_sample_Meta-Llama-3-70B_pretrain_20240512_020339) on the curated_instructions_3k dataset.
19
 
20
- ## Model description
21
 
22
- More information needed
23
 
24
- ## Intended uses & limitations
25
 
26
- More information needed
 
 
27
 
28
- ## Training and evaluation data
29
 
30
- More information needed
31
 
32
- ## Training procedure
33
 
34
- ### Training hyperparameters
35
 
36
- The following hyperparameters were used during training:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37
  - learning_rate: 3e-06
38
  - train_batch_size: 2
39
  - eval_batch_size: 4
@@ -49,13 +67,39 @@ The following hyperparameters were used during training:
49
  - num_epochs: 1.0
50
  - mixed_precision_training: Native AMP
51
 
52
- ### Training results
 
 
 
 
 
 
 
 
 
 
 
53
 
 
54
 
 
 
 
 
 
 
 
 
 
 
55
 
56
- ### Framework versions
57
 
58
- - Transformers 4.40.2
59
- - Pytorch 2.3.0
60
- - Datasets 2.19.1
61
- - Tokenizers 0.19.1
 
 
 
 
 
1
  ---
2
+ license: cc-by-4.0
3
+ datasets:
4
+ - PKU-Alignment/ProgressGym-HistText
5
+ - PKU-Alignment/ProgressGym-TimelessQA
6
+ base_model:
7
+ - PKU-Alignment/ProgressGym-HistLlama3-70B-C019-pretrain
8
+ - meta-llama/Meta-Llama-3-70B
 
 
9
  ---
10
 
11
+ # ProgressGym-HistLlama3-70B-C019-instruct
 
12
 
13
+ ## Overview
14
 
15
+ #### The ProgressGym Framework
16
 
17
+ ![Framework Diagram](./readme-assets/main-diagram.png)
18
 
19
+ **ProgressGym-HistLlama3-70B-C019-instruct** is part of the **ProgressGym** framework for research and experimentation on *progress alignment* - the emulation of moral progress in AI alignment algorithms, as a measure to prevent risks of societal value lock-in.
20
 
21
+ To quote the paper [*ProgressGym: Alignment with a Millennium of Moral Progress*](https://arxiv.org/abs/2406.20087):
22
 
23
+ > Frontier AI systems, including large language models (LLMs), hold increasing influence over the epistemology of human users. Such influence can reinforce prevailing societal values, potentially contributing to the lock-in of misguided moral beliefs and, consequently, the perpetuation of problematic moral practices on a broad scale.
24
+ >
25
+ > We introduce *progress alignment* as a technical solution to mitigate this imminent risk. Progress alignment algorithms learn to emulate the mechanics of human moral progress, thereby addressing the susceptibility of existing alignment methods to contemporary moral blindspots.
26
 
27
+ #### ProgressGym-HistLlama3-70B-C019-instruct
28
 
29
+ ProgressGym-HistLlama3-70B-C019-instruct is one of the **36 historical language models** in the ProgressGym framework.
30
 
31
+ **ProgressGym-HistLlama3-70B-C019-instruct is under continual iteration.** Improving upon the current version, new versions of the model are currently being trained to reflect historical moral tendencies in ever more comprehensive ways.
32
 
33
+ **ProgressGym-HistLlama3-70B-C019-instruct is a 19th-century historical language model.** Based on [Meta-Llama-3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B), It is continued-pretrained on the 19th-century text data from [ProgressGym-HistText](https://huggingface.co/datasets/PKU-Alignment/ProgressGym-HistText), using the following hyperparameters:
34
 
35
+ - learning_rate: 3e-06
36
+ - train_batch_size: 1
37
+ - eval_batch_size: 2
38
+ - seed: 42
39
+ - distributed_type: multi-GPU
40
+ - num_devices: 32
41
+ - gradient_accumulation_steps: 4
42
+ - total_train_batch_size: 128
43
+ - total_eval_batch_size: 64
44
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
45
+ - lr_scheduler_type: polynomial
46
+ - lr_scheduler_warmup_ratio: 0.075
47
+ - num_epochs: 0.5
48
+ - mixed_precision_training: Native AMP
49
+
50
+ ... where training results can be found in `all_results.json`, `trainer_log.jsonl`, and `training_loss.png` of the pretrain model.
51
+
52
+ Note that the training data volume for the continued pretraining stage is capped at 300MB. When the corresponding century's corpus exceeds this volume, the training data is randomly sampled to fit the volume.
53
+
54
+ **ProgressGym-HistLlama3-70B-C019-instruct is an instruction-tuned language model.** It is tuned on [ProgressGym-TimelessQA](https://huggingface.co/datasets/PKU-Alignment/ProgressGym-TimelessQA), using the following hyperparameters:
55
  - learning_rate: 3e-06
56
  - train_batch_size: 2
57
  - eval_batch_size: 4
 
67
  - num_epochs: 1.0
68
  - mixed_precision_training: Native AMP
69
 
70
+ ... where training results can be found in `all_results.json`, `trainer_log.jsonl`, and `training_loss.png` of the instruct model.
71
+
72
+
73
+ ## Links
74
+
75
+ - **[Paper Preprint]** [ProgressGym: Alignment with a Millennium of Moral Progress](https://arxiv.org/abs/2406.20087)
76
+ - **[Leaderboard & Interactive Playground]** PKU-Alignment/ProgressGym-LeaderBoard *(coming soon)*
77
+ - **[Github Codebase]** PKU-Alignment/ProgressGym *(coming soon)*
78
+ - **[Huggingface Data & Model Collection]** [PKU-Alignment/ProgressGym](https://huggingface.co/collections/PKU-Alignment/progressgym-666735fcf3e4efa276226eaa)
79
+ - **[PyPI Package]** *(coming soon)*
80
+
81
+ ## Citation
82
 
83
+ If the datasets, models, or framework of ProgressGym help you in your project, please cite ProgressGym using the bibtex entry below.
84
 
85
+ ```text
86
+ @article{progressgym,
87
+ title={ProgressGym: Alignment with a Millennium of Moral Progress},
88
+ author={Tianyi Qiu and Yang Zhang and Xuchuan Huang and Jasmine Xinze Li and Jiaming Ji and Yaodong Yang},
89
+ journal={arXiv preprint arXiv:2406.20087},
90
+ eprint={2406.20087},
91
+ eprinttype = {arXiv},
92
+ year={2024}
93
+ }
94
+ ```
95
 
96
+ ## Ethics Statement
97
 
98
+ - **Copyright information of historical text data sources**:
99
+ - Project Gutenberg, one among our four source of our historical text data, consists only of texts in the public domain.
100
+ - For the text that we draw from Internet Archive, we only include those that uploaded by *Library of Congress*, which are texts freely released online by the U.S. Library of Congress for research and public use.
101
+ - The text data from Early English Books Online are, according to their publisher, "freely available to the public" and "available for access, distribution, use, or reuse by anyone".
102
+ - The last remaining source of our historical text data, the Pile of Law dataset, is released under a Creative Commons license, which we adhere to in our use.
103
+ - **Reproducibility**: To ensure reproducibility, we open-source all the code involved in the production of our main results (including the entire pipeline starting from data collection and model training), as well as the supporting infrastructure (the ProgressGym framework), making replication as easy as running a few simple script files.
104
+ - **Misuse Prevention**: In order to prevent potential misuse of progress alignment algorithms, we have carefully formulated progress alignment as strictly value-neutral, without *a priori* assumptions on the direction of progress. In the event of potential misuse of our dataset, we condemn any misuse attempt to the strongest degree possible, and will work with the research community on whistleblowing for such attempts.
105
+ - **Open-Sourcing**: We confirm that our code, data, and models are to be open-sourced under a CC-BY 4.0 license. We will continue to maintain and update our open-source repositories and models.