This repository contains the official implementation of Spectrum-Guided Task Arithmetic, a project for the Deep Learning and Artificial Intelligence (DLAI) 2025/2026 course at Sapienza University of Rome.
Model merging via Task Arithmetic offers a zero-shot, computationally efficient alternative to multi-task fine-tuning by linearly combining the task vectors of fine-tuned networks. However, merging models fine-tuned on orthogonal tasks (e.g., highly syntactic logic tasks vs. highly semantic emotion classification) often leads to catastrophic weight interference.
Drawing on the architectural feature hierarchy of Transformers—where lower layers process general syntax/linguistic structures and upper layers encode task-specific semantics—we propose a Layer-Wise Merging strategy:
Our empirical results demonstrate that this bottom-up layer-wise architecture effectively mitigates interference, significantly outperforming global full-model merging.
evaluate_dataset.py: Baseline script implementing global full-model merging (merging L0–L11).evaluate_hybrid_final.py: Core script implementing our proposed bottom-up merging (L0–L5 & embeddings merged, L6–L11 pure).evaluate_hybrid_reverse.py & evaluate_reverse.py: Ablation studies implementing reverse layer-wise merging (L0–L5 pure, L6–L11 merged) to validate our structural hierarchy hypothesis.test_vram.py: Utility script to inspect GPU hardware and VRAM consumption before running pipelines.DLAI_report.pdf: The final, fully compiled 2-page project report containing theoretical formalizations and results.report/main.tex: The raw LaTeX source code for the project report.This project leverages the modern and fast Python package manager uv.
git clone https://github.com/Alby7503/dlai-task-arithmetic.git
cd dlai-task-arithmetic
uv sync
.venv and installs PyTorch (with CUDA 12.1 support), Transformers, Datasets, and Scikit-Learn.All pipelines can be executed cleanly on any standard Windows, Linux, or macOS terminal using uv:
uv run test_vram.py
uv run evaluate_dataset.py
uv run evaluate_hybrid_final.py
uv run evaluate_hybrid_reverse.py
Experiments were validated on an NVIDIA GeForce GTX 1080 GPU (8.00 GB VRAM) on the SNLI (NLI) and GoEmotions (Emotion classification) test datasets:
| Strategy | Layer Merging | NLI Acc. (SNLI) | Emotion Macro F1 (GoEmotions) |
|---|---|---|---|
| Pure Task A (Emotions) | None | - | 52.10 |
| Pure Task B (NLI) | None | 89.50% | - |
| Full Merge (Baseline) | L0–L11 | 79.39% | 38.20 |
| Bottom-Up (Proposed) | L0–L5 & Embeddings | 88.35% | 48.03 |
| Reverse (Ablation) | L6–L11 | 88.35% | 45.03 |
[!NOTE] The ablation study shows a 3.0 point collapse in Emotion classification F1-score when interference is introduced into the top semantic layers (48.03 $\rightarrow$ 45.03), while logical accuracy remains stable (88.35%). This empirically proves our hypothesis that upper-layer specialized parameters are highly fragile and must be preserved task-pure.
This repository contains the official implementation of Spectrum-Guided Task Arithmetic, a project for the Deep Learning and Artificial Intelligence (DLAI) 2025/2026 course at Sapienza University of Rome.
Model merging via Task Arithmetic offers a zero-shot, computationally efficient alternative to multi-task fine-tuning by linearly combining the task vectors of fine-tuned networks. However, merging models fine-tuned on orthogonal tasks (e.g., highly syntactic logic tasks vs. highly semantic emotion classification) often leads to catastrophic weight interference.
Drawing on the architectural feature hierarchy of Transformers—where lower layers process general syntax/linguistic structures and upper layers encode task-specific semantics—we propose a Layer-Wise Merging strategy:
Our empirical results demonstrate that this bottom-up layer-wise architecture effectively mitigates interference, significantly outperforming global full-model merging.
evaluate_dataset.py: Baseline script implementing global full-model merging (merging L0–L11).evaluate_hybrid_final.py: Core script implementing our proposed bottom-up merging (L0–L5 & embeddings merged, L6–L11 pure).evaluate_hybrid_reverse.py & evaluate_reverse.py: Ablation studies implementing reverse layer-wise merging (L0–L5 pure, L6–L11 merged) to validate our structural hierarchy hypothesis.test_vram.py: Utility script to inspect GPU hardware and VRAM consumption before running pipelines.DLAI_report.pdf: The final, fully compiled 2-page project report containing theoretical formalizations and results.report/main.tex: The raw LaTeX source code for the project report.This project leverages the modern and fast Python package manager uv.
git clone https://github.com/Alby7503/dlai-task-arithmetic.git
cd dlai-task-arithmetic
uv sync
.venv and installs PyTorch (with CUDA 12.1 support), Transformers, Datasets, and Scikit-Learn.All pipelines can be executed cleanly on any standard Windows, Linux, or macOS terminal using uv:
uv run test_vram.py
uv run evaluate_dataset.py
uv run evaluate_hybrid_final.py
uv run evaluate_hybrid_reverse.py
Experiments were validated on an NVIDIA GeForce GTX 1080 GPU (8.00 GB VRAM) on the SNLI (NLI) and GoEmotions (Emotion classification) test datasets:
| Strategy | Layer Merging | NLI Acc. (SNLI) | Emotion Macro F1 (GoEmotions) |
|---|---|---|---|
| Pure Task A (Emotions) | None | - | 52.10 |
| Pure Task B (NLI) | None | 89.50% | - |
| Full Merge (Baseline) | L0–L11 | 79.39% | 38.20 |
| Bottom-Up (Proposed) | L0–L5 & Embeddings | 88.35% | 48.03 |
| Reverse (Ablation) | L6–L11 | 88.35% | 45.03 |
[!NOTE] The ablation study shows a 3.0 point collapse in Emotion classification F1-score when interference is introduced into the top semantic layers (48.03 $\rightarrow$ 45.03), while logical accuracy remains stable (88.35%). This empirically proves our hypothesis that upper-layer specialized parameters are highly fragile and must be preserved task-pure.