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..
Lipsum-FT: Robust Fine-Tuning of Zero-Shot Models Using Random Text Guidance
Authors: Giung Nam, Byeongho Heo, Juho Lee
ICLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on distribution shift scenarios in Domain Net and Image Net confirm the superiority of our proposed Lipsum-FT approach over existing robust fine-tuning methods. |
| Researcher Affiliation | Collaboration | Giung Nam1 Byeongho Heo2 Juho Lee1,3 1KAIST AI 2NAVER AI Lab 3AITRICS |
| Pseudocode | No | No pseudocode or algorithm blocks labeled 'Pseudocode' or 'Algorithm' were found. |
| Open Source Code | Yes | Code is available at https://github.com/cs-giung/lipsum-ft. |
| Open Datasets | Yes | Image Net (IN; Russakovsky et al., 2015)... Domain Net (Peng et al., 2019) |
| Dataset Splits | Yes | We also evaluate fine-tuned models every 1000 iterations and choose the model that demonstrates the highest performance on the reference validation data, which corresponds to Domain Net-R for the Domain Net scenario and Image Net for the Image Net scenario. |
| Hardware Specification | Yes | All experiments are carried out using eight TPUv2 or TPUv3 cores, supported by TPU Research Cloud. |
| Software Dependencies | No | The paper lists software components but does not provide specific version numbers for them. For example, it mentions 'JAX (Bradbury et al., 2018)' but not 'JAX 0.3.25'. |
| Experiment Setup | Yes | Unless specified, the mini-batch size was set to 256, the peak learning rate was configured as 1e-05 with the initial 10% of steps were dedicated to the warm-up phase. |