Research map / Architectures and efficiency / Pruning and efficient tuning
Fine-Tuning of Transformer models with Frames
Harshavardhan Adepu, Li Zhang, Sanjiv Kumar, Vikas Singh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use per-layer random mask instead of shared random mask (ablation) | FrameFT · GLUE avg (RoBERTa-base) | avg score: 86.18 -> 86.26 | No clear effect |
| use gradient-based sparsity mask instead of random (ablation) | FrameFT · GLUE avg (RoBERTa-base) | avg score: 86.18 -> 74.83 | Hurt |
| use magnitude-based mask instead of random (ablation) | FrameFT · GLUE avg (RoBERTa-base) | avg score: 86.18 -> 85.60 | Hurt |
| increase number of non-zero coefficients (ablation) | FrameFT · GLUE (RoBERTa-base) | score | Improved |
| increase subspace dimension rho from 2 to 256 (ablation) | FrameFT · CoLA and MRPC (RoBERTa-base) | score | No clear effect |
| vary scaling factor alpha from 10 to 600 (ablation) | FrameFT · Alpaca Llama-2-7B | avg accuracy: 62.97 to 63.62 | No clear effect |
| replace LoRA with FrameFT (1K coefficients) | Llama-2-7B adapter · inference throughput | tokens/sec: 24.4k -> 37.9k | Improved |
| replace LoRA with FrameFT | Llama-2-7B · Alpaca training time | minutes: 238 -> 225 | Improved |
| replace LoRA with FrameFT | ViT-B · image classification (8 datasets) avg | avg accuracy: 77.58 -> 85.80 | Improved |
| replace LoRA with FrameFT | RoBERTa-base · GLUE avg | avg score: 85.2 -> 86.1 | Improved |
| replace LoRA with FrameFT | Llama-2-13B · Alpaca LM-eval avg | avg accuracy: 66.42 -> 66.11 | Hurt |
| replace LoRA with FrameFT | Llama-3.1-8B · Alpaca LM-eval avg | avg accuracy: 68.56 -> 67.87 | Hurt |
Closest papers by meaning
- Structured Unrestricted-Rank Matrices for Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.91)
- SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors (NeurIPS 2024 · similarity 0.89)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.88)
- Be Confident in What You Know: Bayesian Parameter Efficient Fine-Tuning of Vision Foundation Models (NeurIPS 2024 · similarity 0.88)
- From PEFT to DEFT: Parameter Efficient Finetuning for Reducing Activation Density in Transformers (AAAI 2025 · similarity 0.88)
- WeGeFT: Weight‑Generative Fine‑Tuning for Multi‑Faceted Efficient Adaptation of Large Models (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
