PKU-Alignment/ProgressGym-HistLlama3-8B-C013-instruct-v0.2

ProgressGym-HistLlama3-8B-C013-instruct

Overview

The ProgressGym Framework

ProgressGym-HistLlama3-8B-C013-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.

To quote the paper ProgressGym: Alignment with a Millennium of Moral Progress:

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.

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.

ProgressGym-HistLlama3-8B-C013-instruct

ProgressGym-HistLlama3-8B-C013-instruct is one of the 36 historical language models in the ProgressGym framework.

ProgressGym-HistLlama3-8B-C013-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.

ProgressGym-HistLlama3-8B-C013-instruct is a 13th-century historical language model. Based on Meta-Llama-3-8B, It is continued-pretrained on the 13th-century text data from ProgressGym-HistText, using the following hyperparameters:

• learning_rate: 1.5e-05
• train_batch_size: 8
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: polynomial
• lr_scheduler_warmup_steps: 20
• num_epochs: 4.0
• mixed_precision_training: Native AMP

... with the following training results:

Training Loss	Epoch	Step	Validation Loss
1.7594	0.0149	1	1.7163
1.7333	0.0746	5	1.7008
1.6854	0.1493	10	1.6825
1.6897	0.2239	15	1.6701
1.6656	0.2985	20	1.6651
1.7254	0.3731	25	1.6679
1.7178	0.4478	30	1.6542
1.6656	0.5224	35	1.6459
1.6647	0.5970	40	1.6308
1.6645	0.6716	45	1.6205
1.6151	0.7463	50	1.6129
1.6359	0.8209	55	1.6052
1.5885	0.8955	60	1.5995
1.6142	0.9701	65	1.5943
1.4875	1.0448	70	1.5963
1.3844	1.1194	75	1.6118
1.3555	1.1940	80	1.6069
1.3597	1.2687	85	1.6040
1.3737	1.3433	90	1.6071
1.3492	1.4179	95	1.6074
1.3826	1.4925	100	1.6055
1.3533	1.5672	105	1.6035
1.3611	1.6418	110	1.6023
1.328	1.7164	115	1.6022
1.3443	1.7910	120	1.6026
1.3386	1.8657	125	1.6029
1.3396	1.9403	130	1.6029
1.3573	2.0149	135	1.6029
1.3754	2.0896	140	1.6034
1.3229	2.1642	145	1.6044
1.3194	2.2388	150	1.6055
1.3361	2.3134	155	1.6065
1.3231	2.3881	160	1.6072
1.32	2.4627	165	1.6076
1.3406	2.5373	170	1.6078
1.3184	2.6119	175	1.6079
1.2745	2.6866	180	1.6080
1.3024	2.7612	185	1.6079
1.3243	2.8358	190	1.6079
1.3239	2.9104	195	1.6080
1.3349	2.9851	200	1.6081
1.337	3.0597	205	1.6079
1.3091	3.1343	210	1.6078
1.3266	3.2090	215	1.6079
1.3014	3.2836	220	1.6083
1.3153	3.3582	225	1.6086
1.3192	3.4328	230	1.6090
1.315	3.5075	235	1.6093
1.3047	3.5821	240	1.6093
1.3208	3.6567	245	1.6093
1.362	3.7313	250	1.6093
1.3255	3.8060	255	1.6091
1.2941	3.8806	260	1.6089
1.3254	3.9552	265	1.6086

Note that the training data volume for the continued pretraining stage is capped at 3GB. When the corresponding century's corpus exceeds this volume, the training data is randomly sampled to fit the volume.

ProgressGym-HistLlama3-8B-C013-instruct is an instruction-tuned language model. It is tuned on ProgressGym-TimelessQA, using the following hyperparameters. Note, however, that the snapshot at training step 10 is used for the final model, to minimize erosion of the value tendencies learned during continued pretraining; we qualitatively observe that this snapshot still possesses strong instruction-following capabilities.

• learning_rate: 1.5e-05
• train_batch_size: 8
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: polynomial
• lr_scheduler_warmup_steps: 20
• num_epochs: 4.0
• mixed_precision_training: Native AMP

... with the following training results:

Training Loss	Epoch	Step	Validation Loss
0.9805	0.0208	1	0.9737
0.9446	0.1042	5	0.9455
0.8481	0.2083	10	0.8154
0.7794	0.3125	15	0.8123
0.7798	0.4167	20	0.8411
0.8576	0.5208	25	0.8676
0.8852	0.625	30	0.8673
0.8529	0.7292	35	0.8561
0.8224	0.8333	40	0.8470
0.8536	0.9375	45	0.8378
0.662	1.0417	50	0.8294
0.437	1.1458	55	0.8531
0.4402	1.25	60	0.8569
0.4244	1.3542	65	0.8569
0.4495	1.4583	70	0.8547
0.4689	1.5625	75	0.8494
0.4309	1.6667	80	0.8461
0.4299	1.7708	85	0.8446
0.4461	1.875	90	0.8440
0.4474	1.9792	95	0.8439
0.3614	2.0833	100	0.8445
0.3861	2.1875	105	0.8457
0.3829	2.2917	110	0.8473
0.3764	2.3958	115	0.8488
0.3655	2.5	120	0.8500
0.4243	2.6042	125	0.8511
0.3884	2.7083	130	0.8520
0.3634	2.8125	135	0.8528
0.3846	2.9167	140	0.8537
0.3872	3.0208	145	0.8547
0.3869	3.125	150	0.8558
0.3876	3.2292	155	0.8566
0.3844	3.3333	160	0.8573
0.3535	3.4375	165	0.8579
0.3488	3.5417	170	0.8588
0.3464	3.6458	175	0.8598
0.361	3.75	180	0.8607
0.3674	3.8542	185	0.8612
0.3988	3.9583	190	0.8612

Links

• [Paper Preprint] ProgressGym: Alignment with a Millennium of Moral Progress
• [Leaderboard & Interactive Playground] PKU-Alignment/ProgressGym-LeaderBoard
• [Huggingface Data & Model Collection] PKU-Alignment/ProgressGym
• [Github Codebase] PKU-Alignment/ProgressGym
• [Documentation] ProgressGym Documentation
• [PyPI Package] (coming soon - stay tuned!)

Citation

If the datasets, models, or framework of ProgressGym help you in your project, please cite ProgressGym using the bibtex entry below.

@article{progressgym,
  title={ProgressGym: Alignment with a Millennium of Moral Progress},
  author={Tianyi Qiu and Yang Zhang and Xuchuan Huang and Jasmine Xinze Li and Jiaming Ji and Yaodong Yang},
  journal={arXiv preprint arXiv:2406.20087},
  eprint={2406.20087},
  eprinttype = {arXiv},
  year={2024}
}

Ethics Statement

• Copyright information of historical text data sources:
  • Project Gutenberg, one among our four source of our historical text data, consists only of texts in the public domain.
  • 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.
  • 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".
  • 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.
• 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.
• 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.
• 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.