PKU-Alignment/ProgressGym-HistLlama3-8B-C015-instruct-v0.2

ProgressGym-HistLlama3-8B-C015-instruct

Overview

The ProgressGym Framework

ProgressGym-HistLlama3-8B-C015-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-C015-instruct

ProgressGym-HistLlama3-8B-C015-instruct is one of the 36 historical language models in the ProgressGym framework.

ProgressGym-HistLlama3-8B-C015-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-C015-instruct is a 15th-century historical language model. Based on Meta-Llama-3-8B, It is continued-pretrained on the 15th-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: 3.02
• mixed_precision_training: Native AMP

... with the following training results:

Training Loss	Epoch	Step	Validation Loss
2.6141	0.006494	1	2.6354
2.657	0.032468	5	2.6206
2.6337	0.064935	10	2.5846
2.5268	0.097403	15	2.5516
2.5275	0.129870	20	2.5321
2.5005	0.162338	25	2.5131
2.5339	0.194805	30	2.4961
2.5335	0.227273	35	2.4808
2.4252	0.259740	40	2.4643
2.4445	0.292208	45	2.4518
2.4594	0.324675	50	2.4394
2.4498	0.357143	55	2.4287
2.3821	0.389610	60	2.4184
2.4317	0.422078	65	2.4091
2.3931	0.454545	70	2.4001
2.3695	0.487013	75	2.3934
2.3981	0.519481	80	2.3855
2.3952	0.551948	85	2.3789
2.4137	0.584416	90	2.3721
2.3614	0.616883	95	2.3669
2.3467	0.649351	100	2.3612
2.4012	0.681818	105	2.3569
2.3224	0.714286	110	2.3528
2.3348	0.746753	115	2.3483
2.3573	0.779221	120	2.3448
2.306	0.811688	125	2.3412
2.342	0.844156	130	2.3382
2.3045	0.876623	135	2.3356
2.2959	0.909091	140	2.3330
2.3545	0.941558	145	2.3305
2.3446	0.974026	150	2.3285
2.2502	1.006494	155	2.3268
2.0791	1.038961	160	2.3347
2.1034	1.071429	165	2.3399
2.095	1.103896	170	2.3358
2.0627	1.136364	175	2.3346
2.0408	1.168831	180	2.3357
2.0575	1.201299	185	2.3364
2.0976	1.233766	190	2.3349
2.0668	1.266234	195	2.3336
2.0579	1.298701	200	2.3329
2.0756	1.331169	205	2.3326
2.1174	1.363636	210	2.3325
2.0663	1.396104	215	2.3325
2.0941	1.428571	220	2.3324
2.1074	1.461039	225	2.3324
2.1251	1.493506	230	2.3322
2.0629	1.525974	235	2.3318
2.0872	1.558442	240	2.3312
2.0994	1.590909	245	2.3310
2.0879	1.623377	250	2.3308
2.0623	1.655844	255	2.3305
2.1054	1.688312	260	2.3303
2.0736	1.720779	265	2.3301
2.1146	1.753247	270	2.3300
2.0444	1.785714	275	2.3301
2.0541	1.818182	280	2.3301
2.1333	1.850649	285	2.3300
2.1101	1.883117	290	2.3299
2.0234	1.915584	295	2.3298
2.0671	1.948052	300	2.3298
2.083	1.980519	305	2.3298
2.0417	2.012987	310	2.3299
2.0784	2.045455	315	2.3303
2.058	2.077922	320	2.3308
2.0524	2.110390	325	2.3312
2.0318	2.142857	330	2.3316
2.0914	2.175325	335	2.3318
2.0319	2.207792	340	2.3320
2.0099	2.240260	345	2.3322
2.075	2.272727	350	2.3323
2.0444	2.305195	355	2.3324
2.0428	2.337662	360	2.3325
2.0612	2.370130	365	2.3326
2.1078	2.402597	370	2.3327
2.0643	2.435065	375	2.3327
2.0667	2.467532	380	2.3326
2.0285	2.500000	385	2.3324
2.0571	2.532468	390	2.3322
2.0209	2.564935	395	2.3322
2.0537	2.597403	400	2.3323
2.0138	2.629870	405	2.3324
2.0772	2.662338	410	2.3324
2.039	2.694805	415	2.3323
2.0181	2.727273	420	2.3322
2.0484	2.759740	425	2.3320
2.0224	2.792208	430	2.3320
2.0732	2.824675	435	2.3320
2.0499	2.857143	440	2.3321
2.0498	2.889610	445	2.3321
2.0472	2.922078	450	2.3320
2.1327	2.954545	455	2.3319
2.0642	2.987013	460	2.3319
2.0654	3.019481	465	-

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-C015-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
0.8675	0.1042	5	0.8585
0.8415	0.2083	10	0.8063
0.8225	0.3125	15	0.8210
0.806	0.4167	20	0.8412
0.8139	0.5208	25	0.8702
0.8978	0.625	30	0.8631
0.814	0.7292	35	0.8550
0.7989	0.8333	40	0.8473
0.8769	0.9375	45	0.8383
0.7244	1.0417	50	0.8278
0.4644	1.1458	55	0.8387
0.4488	1.25	60	0.8680
0.3973	1.3542	65	0.8718
0.443	1.4583	70	0.8596
0.4346	1.5625	75	0.8514
0.4701	1.6667	80	0.8461
0.4344	1.7708	85	0.8437
0.4274	1.875	90	0.8434
0.4771	1.9792	95	0.8434
0.3876	2.0833	100	0.8439
0.3698	2.1875	105	0.8451
0.407	2.2917	110	0.8465
0.374	2.3958	115	0.8482
0.3945	2.5	120	0.8498
0.3753	2.6042	125	0.8513
0.3721	2.7083	130	0.8528
0.3718	2.8125	135	0.8542
0.3773	2.9167	140	0.8555
0.3723	3.0208	145	0.8565
0.374	3.125	150	0.8576
0.3728	3.2292	155	0.8588
0.3686	3.3333	160	0.8598
0.3617	3.4375	165	0.8607
0.3546	3.5417	170	0.8613
0.3707	3.6458	175	0.8619
0.3739	3.75	180	0.8625
0.3617	3.8542	185	0.8632
0.3591	3.9583	190	0.8637

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.