Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Scaling Law with Learning Rate Annealing
Authors: Howe Tissue, Venus Wang, Lu Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our findings hold across a range of hyperparameters and model architectures and can extend to scaling effect of model sizes. |
| Researcher Affiliation | Academia | Howe Tissue EMAIL Venus Wang EMAIL Lu Wang EMAIL |
| Pseudocode | No | The paper describes methods and formulas but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | Our experiments are conducted under the standard and popular framework, Megatron, which is a fully open-sourced and reproduciable pre-training library. |
| Open Datasets | Yes | Our datasets include Fineweb (Penedo et al., 2024) and Red Pajama-CC (Computer, 2023). The primary training dataset, Fineweb, is also open-sourced and reproduciable. |
| Dataset Splits | No | Given the same training and validation dataset, the same model size, the same training hyper-parameters such as warmup steps, max learning rate ηmax and batch size, the language modeling loss at training step s empirically follows the equation... Train Dataset Fineweb Val Dataset Red Pajama-CC |
| Hardware Specification | Yes | For compute resource, we use a GPU cluster of A100. The experiments are primarily conducted using 16 8 GPU cards. |
| Software Dependencies | No | We use the standard and popular pre-training library, Megatron,(https://github.com/NVIDIA/ Megatron-LM) as our training framework. |
| Experiment Setup | Yes | In this work, we use multiple sets of experimental setups, in order to validate that our equation can work across different experimental setups. For clarification, we present the experimental setup list as shown in Table 3. |