LLM360: Towards Fully Transparent Open-Source LLMs

Summary

This paper presents a fully open-sourced language modeling project, whose aim is to release all code, training data, intermediate checkpoints and results, and also serves as a technical report for two 7B parameter language models, Amber (trained on 1.3T tokens of English text) and Crystal (trained on 1.5T tokens of English text and code). Along with these pretrained models, three finetuned variants of these models, AmberChat and CrystalChat (SFT) and AmberSafe (DPO) were described and evaluated. Amber and AmberChat were evaluated on ARC, Hellaswag, MMLU, and TruthfulQA, and Crystal was evaluated additionally on HumanEval (from the Codex paper) and MBPP.

Rating

8

Confidence

3

Ethics flag

1

Reasons to accept

The paper presents a strong motivation for fully open language model research, and includes many useful training details. Moreover trained models and the intermediate checkpoints and WandB logs are useful artifacts for the community.

Reasons to reject

The paper could use more detailed evaluations and further discussion of existing results. CrystalChat does not have any evaluations in the paper currently. 1. Chat-focused evaluation (e.g. AlpacaEval) and safety evaluations (e.g. Toxigen) of AmberChat, CrystalChat and AmberSafe would helpful to see. 2. Further discussion of the performance on Amber (the base model) on the OpenLLM leaderboard benchmarks would be informative. What would make the model's performance more comparable to better 7B models? Is it the training data mix or the size? I understand that concrete observations cannot be made here because all the training details of the other models are not known, but a discussion on how the model could be improved would be helpful. 3. The draft currently does not have any details of the carbon footprint of training the models.

Summary

This work present a fully open-source LLMs initiative LLM360, where not just model weights, but all the training data, training code, model checkpoints and intermediate results are made available to the community. This initiative is significant for the community as it addresses several important challenges LLM research is dealing with - data provenance, reproducibility and open collaboration. In the paper, the authors outline the LLM360 approach, as well as pretrain two new LLMs (Amber and Crystal) from scratch and release them under the LLM360 framework, together with all the associated core open-sourced artifacts. The paper is very well written and clearly structured, the rationale behind open-sourcing the core artifacts is properly explained, the process of pre-training newly presented LLMs and their evaluation are both discussed in sufficient details. The take-home messages and potential use cases of LLM360 make this work useful also for the future researchers and practitioners. The authors are transparent about the issues encountered during training and make suggestions about the responsible usage of their models.

Rating

9

Confidence

4

Ethics flag

1

Reasons to accept

- this ambitious work contributes to a more inclusive and collaborative research environment by reinforcing visibility and access to LLM training, fine-tuning, and evaluation processes. - this work has a strong technological impact, as it presents two newly pre-trained from scratch open-source LLMs with the associated open training datasets, data processing code, training code, hyperparameters, configurations, model checkpoints, collected logs and metrics. These high quality resources will enable future high quality and impactful work. - the described pre-training process of two new LLMs presents a strong empirical foundation. It provides useful insights for the community and is beneficial for understanding the details and potential major issues of the pre-training process.

Reasons to reject

I do not see significant reasons to reject this paper. In order to improve it for the camera-ready version, I would suggest adding a couple of recent models to the related work discussion, specifically Aya (Üstün, Ahmet, et al. Aya model: An instruction finetuned open-access multilingual language model", 2024) and StarCoder 2 (Lozhkov, Anton, et al. "StarCoder 2 and The Stack v2: The Next Generation", 2024 ). It would also be beneficial to read a bit about how the authors see the future of LLM360 in terms of the open collaboration with the community, e.g. do they expect external contributions to LLM360 in the future?

Questions to authors

Correct the typo under the Issues Encountered During Pre-training section on page 9: "...which can be a major help ***the*** the community".

Summary

The LLM360 initiative addresses a critical need in the AI research community for complete transparency in the training of Large Language Models (LLMs). The quality of this work is evident in the comprehensive approach to open-sourcing, which includes training code, data, model checkpoints, and intermediate results.

Rating

6

Confidence

5

Ethics flag

1

Reasons to accept

1. LLM360’s full disclosure of training processes, including intermediate checkpoints and data provenance, addresses a significant gap in current LLM research practices. 2. By providing all necessary artifacts for replication, LLM360 significantly enhances the reproducibility of LLM research, aligning with scientific principles.

Reasons to reject

1. The paper lacks extensive comparative analysis with existing LLMs, which could better demonstrate the efficacy and efficiency of the LLM360 models. 2. The long-term sustainability of continually open-sourcing such extensive resources is not discussed, which could be a potential hurdle in widespread adoption.

Questions to authors

1. How do you plan to sustain the LLM360 initiative in terms of funding and resource allocation? 2. Can you elaborate on any legal or ethical considerations in open-sourcing such extensive datasets and model details?

Summary

This paper discusses the importance of open-source large language models and makes a step in this direction by proposing LLM360, an initiative which encourages the community to make publicly available training data, code, model checkpoints and intermediate results. The authors make a step in this direction by releasing two 7B LLMs pretrained from scratch for the English language, Amber (1.3T tokens) and Crystal (1.4T tokens), altogether with their pre-training data, code, intermediate checkpoints and analyses. In addition, the authors also release corresponding fine-tuned models AmberChat, AmberSafe, CrystalChat.

Rating

6

Confidence

4

Ethics flag

2

Reasons to accept

The paper makes an important contribution towards transparent open-source LLMs by releasing to the community two English LLM models pre-trained from scratch along with the training data, code, model checkpoints and performance reports. The work is timely and would greatly benefit the open-source research community. The authors acknowledge issues encountered during pre-training with the goal of being fully transparent about the model training and potential pitfalls.

Reasons to reject

The paper does not discuss the goal of fine-tuning the models. It is unclear from the paper what AmberChat, AmberSafe, CrystalChat have been finetuned to do and for which tasks they can be readily used. The authors do not provide an explanation why on some metrics the Amber model does well, while on some other metrics (ARC) the model is behind the curve. This statement should be clarified: “We also find that our finetuned AMBER models are relatively strong, even compared with other similar models.” The paper frequently mentions the Pythia (Biderman et al., 2023b) model, but there is no direct comparison between the models of this paper with Pythia. No discussion related to data quality used for pre-training and safety aspects of the model. The paper would greatly benefit from more detailed discussion of data selection and analysis of model biases, controllability and trustworthiness. “For future work, we are conducting a more detailed analysis of AMBER and CRYSTAL’s base models as well as their fine-tuned models. Detailed results will be released and discussed in their respective technical reports.” - In my view, this detailed model analysis should not be left to future work, but included in the current paper.

Questions to authors

What guardrails are in place to contain the potential risks associated with the models you plan to release?

Ethics concerns details

The authors plan to release open-source models to the community, however the paper does not include a detailed analysis of Discrimination / Bias / Fairness Concerns associated with these models.