Research map / Continual learning / Domain and prompt-based
DoFIT: Domain-aware Federated Instruction Tuning with Alleviated Catastrophic Forgetting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DoFIT, a domain-aware federated instruction tuning framework that aggregates overlapping top-k LoRA modules across domains on an inter-domain server and initializes intra-domain weights with a proximal perturbation to reduce domain-information catastrophic forgetting.
Problem
Conventional federated instruction tuning (FIT) treats client-aware heterogeneity only and ignores domain-aware heterogeneity, so when a scarce domain is supplemented with a related general domain, domain information is catastrophically forgotten and per-domain performance drops.
Approach
Two-level server design: intra-domain servers upload only top-k LoRA modules ranked by squared norm, the inter-domain server averages only overlapping modules, and intra-domain initialization adds a scaled proximal perturbation (projected inter-domain difference) instead of overwriting the original modules.
Key results
- On the finance domain with the Finance&General mix, DoFITα=1.0 reaches Avg:4 accuracy 0.849 and F1 0.858 versus 0.801 / 0.813 for conventional FIT32qv.
- On the Medical&General mix, DoFITα=1.2 raises MedQA accuracy to 0.261 from 0.174 for FIT32qv (+0.094 relative to the paired baseline in the table) and average to 0.258 from 0.196.
- Conventional FIT trained on Finance&General performs worse than Finance-only training, while the top-15 DoFIT-base variant surpasses FIT on the mixed setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use domain-aware top-k aggregation rather than conventional cross-domain FedAvg | FIT on Finance&General · Finance&General, Avg:4 | Acc: 0.761 -> 0.791 | Helped |
| Add proximal perturbation initialization to the top-k aggregation baseline | Basetop15 on Finance&General · Finance&General, Avg:4 | Acc: 0.791 -> 0.797 | Helped |
| Add proximal perturbation initialization to the top-k aggregation baseline | Basetop15 on Finance&General · Finance&General, Avg:4 | F1: 0.798 -> 0.804 | Helped |
| Use top-15 rather than top-10 important modules (ablation) | DoFIT-base on Finance&General · Finance&General, Avg:4 | Acc: 0.771 -> 0.791 | Helped |
| Use top-20 rather than top-15 important modules (ablation) | DoFIT-base on Finance&General · Finance&General, Avg:4 | Acc: 0.791 -> 0.783 | Hurt |
| Set proximal perturbation scaling factor α to 0.5 rather than 1.0 (ablation) | DoFIT on Finance&General · Finance&General, Avg:4 | Acc: 0.797 -> 0.797 | No clear effect |
| Set proximal perturbation scaling factor α to 1.5 rather than 1.0 (ablation) | DoFIT on Finance&General · Finance&General, Avg:4 | Acc: 0.797 -> 0.785 | Hurt |
| Add top-30 domain-aware aggregation rather than conventional FIT | FIT on Medical&General · Medical&General, Avg | accuracy: 0.196 -> 0.205 | Helped |
| Add proximal perturbation initialization to top-30 aggregation | Basetop30 on Medical&General · Medical&General, Avg | accuracy: 0.205 -> 0.258 | Helped |
| Rank modules by gradient square norm instead of LoRA weight square norm (ablation) | DoFIT on Finance · Finance, Avg:4 | Acc: 0.797 -> 0.802 | Helped |
| Rank modules by sum of singular values instead of LoRA weight square norm (ablation) | DoFIT on Finance · Finance, Avg:4 | Acc: 0.797 -> 0.789 | Hurt |
| Select top-k modules using gradient/singular-value domain-gap criteria (ablation) | DoFIT on Finance · Finance, Avg:4 | Acc: 0.797 -> 0.401 | Hurt |
Surprising findings
- Adding general-domain data (Alpaca-GPT4) to finance training lowers conventional FIT performance below finance-only training.
- Aggregating modules selected by gradient or singular-value criteria with smaller cross-domain gaps performs worse than the simple squared-norm top-k selection.
Limitations
- Only two domains are studied (one specific domain plus General), so multi-domain (more than two) settings are untested.
- Evaluation is limited to a few classification and medical QA benchmarks with a single frozen Llama2-7B backbone.
- The method relies on a heuristic squared-norm top-k module selection and a scaling factor α that needs tuning.
Experimental setup
Datasets: FinGPT Alpaca-GPT4 MedAlpaca FPB FiQA-SA TFNS NWGI MedQA MedMCQA
Benchmarks: FPB FiQA-SA TFNS NWGI MedQA MedMCQA
Metrics: accuracy F1
Baselines: Local training FIT32qv FIT16qvd FIT32d DoFIT-base FedGP FedGP-g FPL GPT-3.5 GPT-4 FedProx
Builds on: FedAvg LoRA FIT (Ye et al.) orthogonal learning for continual learning
Closest papers by meaning
- Federated Orthogonal Training: Mitigating Global Catastrophic Forgetting in Continual Federated Learning (ICLR 2024 · similarity 0.85)
- Large Continual Instruction Assistant (ICML 2025 · similarity 0.84)
- The Best Instruction-Tuning Data are Those That Fit (NeurIPS 2025 · similarity 0.84)
- Mix-CPT: A Domain Adaptation Framework via Decoupling Knowledge Learning and Format Alignment (ICLR 2025 · similarity 0.84)
- Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.84)
- CAN: Leveraging Clients As Navigators for Generative Replay in Federated Continual Learning (ICML 2025 · similarity 0.84)
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.