The paper investigates how the output format of instruction‑tuned models confounds assessments of data quality and model capability. By analyzing gradient signatures across multiple tasks, interfaces, and model families, the authors demonstrate that the interface can mask or exaggerate performance signals, with certain skills appearing only under specific output formats. They show that the interface‑varying residual is not noise but captures task‑specific information, and that correcting a single generation budget can flip the measured effect of fine‑tuning on a benchmark. "whyItMatters":"The study highlights that current evaluation practices may misrepresent a model’s true capabilities because they rely on the output interface rather than the underlying content, potentially leading to misleading conclusions about data quality and model performance."
arXiv:2610.33969v1 Announce Type: new Abstract: "Data quality" is widely invoked in instruction tuning, but the term is often underspecified. We study it as an operational property by asking which si...
arXiv:2609.01244v1 Announce Type: new Abstract: Every supervised fine-tuning run forces the same chain of decisions, such as learning rate, batch size, LoRA or full fine-tuning, how many epochs, whic...
The paper introduces CoDIT, a contrastive decoding technique that separates instruction-following behavior from pre-trained world knowledge in large language models. By generating responses that emphasize post-training instruction capabilities while suppressing shared pre-trained knowledge, CoDIT creates instruction-tuning datasets that lead to consistently better model performance than directly generated responses or existing public datasets. The authors also provide theoretical and empirical evidence that CoDIT effectively distills instruction-tuning knowledge from model parameters into text, facilitating cross-architecture transfer.
Every supervised fine-tuning run forces the same chain of decisions, such as learning rate, batch size, LoRA or full fine-tuning, how many epochs, which optimiser, and what data to feed the model. Eac...
arXiv:2607. 15299v1 Announce Type: cross Abstract: In this paper, we propose MLLM-DataEngine, a novel closed-loop system that bridges data generation, model training, and evaluation.
arXiv:2605.07111v3 Announce Type: replace-cross Abstract: Recent literature on fine-tuning Large Language Models highlights a fundamental debate. While Full Fine-Tuning (FFT) provides greater represe...
arXiv:2504. 15610v4 Announce Type: replace Abstract: Fine-tuning a 7B language model for specialized advising is attractive in resource-constrained settings, but multi-epoch runs routinely exceed the wall-clock limits of the free-tier GPUs (Kaggle, Colab) such users rely on.
arXiv:2606. 17660v1 Announce Type: cross Abstract: Fine-tuning large language models (LLMs) is compute-intensive and error-prone: model performance depends sensitively on data quality and hyperparameter choices, and na\"ive runs can even degrade model performance.
The paper introduces SSR, a simple yet effective data selection strategy for million‑scale multi‑modal instruction‑tuning. SSR uses interaction styles as a proxy for structural diversity and decomposes data quality into capability‑specific scores, forming a rich profile for each sample. By casting selection as submodular coverage maximization, SSR consistently outperforms random sampling across two architecture stacks, multiple data pools, and 14 benchmarks, achieving 99.1% of full‑data performance with only 30% of a 2.6M pool.
arXiv:2610.07553v1 Announce Type: cross Abstract: LoRA fine-tuning adapts small language models (SLMs) to heterogeneous instruction data within a low-rank update subspace, making it vulnerable to thr...
The paper investigates how supervised fine‑tuning (SFT) affects different layers of language models, finding that middle layers (20–80%) remain stable while the final layers are highly sensitive to changes. Using information‑theoretic, geometric, and optimization metrics across 1B‑32B models, the authors identify a depth‑dependent pattern and introduce Mid‑Block Efficient Tuning, which updates only the critical intermediate layers. Experiments show this method outperforms standard LoRA by up to 10.2 % on GSM8K (OLMo2‑7B) with lower parameter overhead, suggesting that alignment can be achieved with localized architectural changes.
arXiv:2604. 05336v2 Announce Type: replace Abstract: Models often fail to complete agentic tasks because they lack core capabilities required by the target environment.