LLM360: Towards Fully Transparent Open-Source LLMs

Thank you for the thoughtful review of our paper, and for the helpful feedback. We respond back to individual comments below.

CrystalChat evaluations

Thanks for pointing this out (this was an accidental omission). We have a full suite of evaluations for CrystalChat, which we will include in our paper, including the Open LLM Leaderboard metrics suite (e.g. ARC-C, GSM8K, HellaSwag, MMLU, TruthfulQA, Winogrande) and coding evaluations (HumanEval, MBPP). For reference, CrystalChat performs on par with Llama-2-7b-Chat on the language evaluations, and on par with CodeLlama-7b-Instruct for the code evaluations—e.g. for code evaluations, CrystalChat scores: HumanEval (Pass@1): 34.12, & MBPP (Pass@1): 39.11.

Chat-focused evaluations

We also have completed chat-focused evaluations, specifically the MT-Bench evaluation. For reference, for a 3-round average on MT-Bench, Amber scores 2.49, AmberChat scores 4.95, and AmberSafe scores 4.97. We will describe this evaluation and include all results in our paper.

Safety evaluations

We have conducted an evaluation on safety and toxicity of our model, which we will add to our paper. This analysis includes evaluations on BOLD [1] (AmberChat: 0.51, AmberSafe: 0.64, Llama2: 0.46; higher is better sentiment) & ToxiGen [2] (AmberChat: 10.8, AmberSafe: 2.0, Llama2: 21.28; higher is more toxicity).

[1] https://arxiv.org/abs/2101.11718 [2] https://arxiv.org/abs/2203.09509

Discussion of performance metrics

We have collected some evidence explaining Amber’s performance, and ways to improve it. One important aspect is simply the data. Although our compute budget did not allow for extensive ablations on data composition, we found (in subsequent training, such as in Crystal) that modifying the data composition, size, and quality, produced a significant impact on Open LLM Leaderboard metrics. Another potential explanation is due to a precision issue during training. As described in appendix E, in an initial phase of our training, we used a suboptimal model precision for some of our checkpoints; we believe that this may have also led to some drop in accuracy. We will add these discussions to our paper.

Carbon footprint of models

Thank you for the idea! We will complete an estimate of the carbon footprint of our models, and add these to the paper. For example, for our Amber model, we estimate that the power consumption is 63.9 MWh, and our carbon footprint is 24.6 tCO2eq.

Thank you for the thoughtful review of our paper, and for the helpful feedback. We respond back to individual comments below.

Additions to related work

Thanks for the additional references to recent models for the related work discussion. We have updated our paper to include these new citations.

Future of LLM360 and open collaboration

We are indeed very enthusiastic about expanding the breadth of the LLM360 project, and have been actively working to increase the number of collaborators (and collaborating organizations) in this project. As a fully-open community, we welcome anyone interested to join our collaboration, and are working on processes and governance that can scale up to greater numbers of members.

Typo

Thank you as well for pointing out this typo. We have fixed this in our paper!

Thank you for the thoughtful review of our paper, and for the helpful feedback. We respond back to individual comments below.

Goals of fine-tuning

We agree with your comment, and will add in a description of our goals in fine-tuning. In short, for AmberChat and CrystalChat, our goal was to produce an instruction-following model that could serve as a chatbot. For AmberSafe, our goal was to carry out a stage of safety-alignment in addition to instructing-tuning, so that the effects of each stage can be studied. We will discuss full details of the fine-tuning procedure and datasets so that our process and aims are fully clear.

Discussion of performance metrics

We have made efforts to understand some of the reasons behind Amber performance, and will include an explanation in our paper. As two examples: we found that improving the data composition and size produced a significant increase in Open LLM Leaderboard metrics where Amber underperformed. Additionally, we found an issue with an incorrect model precision during the initial phase of training Amber (described in Appendix E); once fixed, we found that this also improved model performance on our evaluations.

Comparison with Pythia

We have completed a full comparison of our model against Pythia (and other popular LLMs) on over a dozen open benchmarks, and will add this comparison between models to the paper.

Additional analysis

We agree with your point about a detailed model analysis in future work, and as you suggest will aim to include the full model analysis within the current paper.

Ethics, and discussion of model safety/bias

Thank you for bringing up this important point about ethics considerations. We have conducted an ethics and toxicity evaluation in our model, which we will add to our paper. This analysis includes evaluations on BOLD [1] & ToxiGen [2]. We will also include a discussion on the relevant processes we followed to select training data, and steps we took to assess safety, bias and fairness.

As there exist open-weight models (e.g. Llama 3) which offer stronger capabilities but at a lower level of transparency, we believe that releasing our current models does not add additional major risks to the field. However, in the future, if our trained models become more performant, we do believe that guardrails could be necessary to reduce capabilities before release.

[1] https://arxiv.org/abs/2101.11718

[2] https://arxiv.org/abs/2203.09509

Thank you for the thoughtful review of our paper, and for the helpful feedback. We respond back to individual comments below.

Comparative analysis with existing LLMs

We have conducted broader evaluations of Amber and Crystal, and we will add these results to the final paper, along with a comparative analysis with results reported by other existing LLMs. Specifically, these evaluations include the Open LLM Leaderboard metrics suite (e.g. ARC-C, GSM8K, HellaSwag, MMLU, Winogrande, etc), additional evaluations (BoolQ, DROP, Openbook QA, PIQA), HumanEval metrics, and MBPP metrics.

Discussion of long-term sustainability

We also agree on the importance of sustainability of this project. In the near-term, we are engaging a broader community beyond the LLM360 team, and have secured resources for several newer models. In the longer term, our vision for LLM360 is to promote transparency and reproducibility of LLMs to be adopted by many others. We have seen early signs of such adoption by other projects that used our model checkpoints, data, and evaluation framework, and we hope that continued progress on these fronts will lead to longer term continuation of open-source AI in general. We also see other recent efforts that are investing resources for fully open-source LLM training, such as OLMo, and MAP-Neo, so we are optimistic about the community’s collective interests in sustaining these efforts.

Question about legal/ethical considerations

We respect the risks associated with releasing open-source models; however, we believe that the final models that we release do not add additional major risks, especially since there exist open-weight models (such as Llama 3) which offer stronger capabilities but at a lower level of transparency. We hope that open-source releases of LLMs will enable researchers to study the security and safety issues associated with models of this scale.

As for legality, we aim to follow all best-practices for licensing of our model. Further, we only leverage existing open-source data, and simply release the exact data sequence (associated with model checkpoints) and our data processing procedures. Indeed, we feel that being fully-transparent about the data used in an LLM is the best way to ensure that an organization is adhering to legal practices.

Thanks

We hope that we were able to address each of your comments. Please let us know if there is anything else we could answer which would improve your score.