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..
Skrull: Towards Efficient Long Context Fine-tuning through Dynamic Data Scheduling
Authors: Hongtao Xu, Wenting Shen, Yuanxin Wei, Ang Wang, Guo Runfan, Tianxing Wang, Yong Li, Mingzhen Li, Weile Jia
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that Skrull outperforms Deep Speed by 3.76x on average (up to 7.54x) in real-world long-SFT scenarios. |
| Researcher Affiliation | Collaboration | Hongtao Xu1,2,3 Wenting Shen3 Yuanxin Wei4 Ang Wang3 Guo Runfan2 Tianxing Wang3 Yong Li3 Mingzhen Li2 Weile Jia2 1School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences 2State Key Lab of Processors, Institute of Computing Technology, CAS 3Alibaba Group 4Sun Yat-sen University |
| Pseudocode | Yes | Algorithm 1 Heuristic scheduling algorithm of DACP Algorithm 2 Heuristic Scheduling Algorithm of GDS |
| Open Source Code | No | Justification: We use public data in our experiments. Due to some approval process, we will make our code public as soon as possible. |
| Open Datasets | Yes | As shown in Figure 1a, we observe pronounced variance in the sequence length distribution across real-world Long-SFT datasets, including Wikipedia [2], LMsys Chat1M [25] and Chat QA2-Long-SFT [1]. |
| Dataset Splits | No | All the experiments share the same training settings with <DP=4, CP=8, Batch Size=64>, zero-2 enabled and selective recomputation strategy except for training Qwen-2.5-7B with Chat QA2-long-SFT dataset. Due to the increased memory requirements, we adjust its parallel settings with <DP=2, CP=16, Batch Size=40>. The global batch size is equal to DP size multiplied by Batch Size. |
| Hardware Specification | Yes | We conduct experiments using a testbed consisting of 4 nodes interconnected via a high-performance Infini Band network, with each node equipped with 8 Nvidia H100 GPUs connected via 900GB/s NVLink. |
| Software Dependencies | No | Then, We implement Skrull on top of Deep Speed, a state-of-the-art distributed LLM training system and enable Zero-2 optimization as our baseline. |
| Experiment Setup | Yes | All the experiments share the same training settings with <DP=4, CP=8, Batch Size=64>, zero-2 enabled and selective recomputation strategy except for training Qwen-2.5-7B with Chat QA2-long-SFT dataset. Due to the increased memory requirements, we adjust its parallel settings with <DP=2, CP=16, Batch Size=40>. The global batch size is equal to DP size multiplied by Batch Size. ... Through offline profiling, we configure the Bucket Size to 26K and 13K for Qwen2.5-0.5B and Qwen2.5-7B, respectively. |