Quadrant-based Tuning in LLM Fine-Tuning

• Quadrant-based Tuning is a data selection framework that evaluates training samples using error and uncertainty metrics to partition data into four regimes.
• It employs a two-stage strategy with sample-level triage and token-level pruning to retain valuable examples while discarding redundant or noisy data.
• The approach delivers notable efficiency gains, achieving a +38% performance improvement on SmolLM2-1.7B using only 12.5% of the original data.

Quadrant-based Tuning (Q-Tuning) is a unified framework for maximizing data efficiency during supervised fine-tuning (SFT) of LLMs under computational and budgetary constraints. Unlike conventional approaches that apply sample-level or token-level pruning in isolation, Q-Tuning integrates both dimensions via a diagnostic framework known as the Error–Uncertainty (EU) Plane. By distinguishing the heterogeneous utility of training data across both samples and tokens, Q-Tuning enables dynamic and adaptive data curation that systematically outperforms full-data SFT across multiple benchmarks, achieving, for example, a +38% average improvement using only 12.5% of original data on SmolLM2-1.7B (Wang et al., 28 Sep 2025).

1. The Error–Uncertainty Plane

The foundation of Q-Tuning is the characterization of each training sample by two orthogonal dimensions: error and uncertainty. Error is quantified by the model’s observed per-sample perplexity (PPL), with higher values indicating that the sample is either difficult or targeted by a significant misconception. Uncertainty is measured as the predictive entropy of the model's token-level predictions, where higher entropy reveals greater indecision in the model's output distribution.

Mapping each sample onto this two-dimensional EU Plane allows for an explicit partition into four quadrants, each corresponding to a qualitatively distinct data regime:

Quadrant	Description	Action in Q-Tuning
Q1	Harmful Noise	Discarded
Q2	Valuable Misconception	Retained, apply token pruning
Q3	Redundant Knowledge	Discarded
Q4	Calibration Data	Retained in full

• Q1 (Harmful Noise): High error and high uncertainty; likely to contain mislabeled or noisy data.
• Q2 (Valuable Misconception): High error, low uncertainty; the model is confidently wrong, offering strong correctional signals.
• Q3 (Redundant Knowledge): Low error, low uncertainty; already mastered, offering minimal additional value.
• Q4 (Calibration Data): Low error, high uncertainty; challenging but reliable, essential for model calibration.

This quadrant structuring uniquely enables Q-Tuning’s coordinated sample and token selection.

2. Hierarchical Two-Stage Pruning Strategy

Q-Tuning operationalizes data selection via a two-stage hierarchical approach.

Stage 1: Sample-Level Triage

• For each mini-batch, compute PPL and entropy for all samples.
• Use a bisection search over quantiles (parameters $\alpha$, $\beta$) to adaptively set thresholds for partitioning the EU Plane.
• Retain all Q2 (valuable misconception) and Q4 (calibration) samples, discard Q1 (harmful noise) and Q3 (redundant knowledge).
• When the retained set falls below a target ratio, employ a “supporters” score—normalizing PPL minus entropy—to select additional samples.

Stage 2: Asymmetric Token-Level Pruning

• Within Q2 samples, not all tokens are of equal value; detrimental or confusing tokens are likely present.
• Apply token-level pruning to Q2 samples via a context-aware scoring mechanism (see Section 3).
• Q4 samples are retained in their entirety to preserve the informational structure necessary for proper uncertainty estimation.

This two-stage process ensures both inter-sample and intra-sample redundancy are minimized.

3. Context-Aware Token Importance Scoring

For Q2 samples designated for token pruning, Q-Tuning computes a “smoothed” per-token importance score that incorporates local context:

$$s_i(x, y; f_\theta) = (1 - \lambda) \cdot \text{PPL}_i(x, y; f_\theta) + \lambda \cdot \left( \text{PPL}_{i-1}(x, y; f_\theta) + \text{PPL}_{i+1}(x, y; f_\theta) \right)$$

where $s_i$ is the score at token position $i$, $\lambda \in [0,1]$ determines the weight attributed to neighboring tokens, and $f_\theta$ denotes the model. This smoothed scoring prevents pruning due to isolated PPL spikes and favors contextually-informed pruning.

Tokens within each Q2 sequence are ranked by $s_i$. Only a top $r_{token}$ fraction (lowest scores, i.e., least detrimental) is retained, pruning the remainder.

4. Empirical Gains and Comparative Benchmarks

Q-Tuning demonstrates state-of-the-art efficiency and effectiveness on diverse tasks, including reasoning, reading comprehension, and general instruction following. Key empirical findings include:

• On SmolLM2-1.7B, an average improvement of +38% over full-data SFT is achieved using only 12.5% of the original data.
• Across five benchmarks, Q-Tuning consistently outperforms not only full-data SFT but also fragmented (sample-only or token-only) pruning baselines.
• The unified strategy of joint sample and token pruning eliminates the inefficiencies observed when each dimension is optimized in isolation.

This suggests that coordinated, context-sensitive pruning is essential for maximizing supervised fine-tuning efficacy and resource utilization.

5. Scalability and Computational Considerations

Q-Tuning is optimized for practical application in large-scale settings:

• Computational Efficiency: Removal of uninformative samples and compression of sample content sharply reduces effective training set size, diminishing per-step computational requirements.
• Adaptive Operation: Both sample and token thresholds are dynamically updated at each mini-batch, enabling Q-Tuning to follow the model’s evolving state.
• Minimal Overhead: The bisection search partitions the EU Plane in fewer than ten iterations, adding negligible computational overhead.
• Data Utilization: By concentrating learning on maximally informative regions of the data distribution, Q-Tuning maintains or improves performance under strict data and compute budgets.

A plausible implication is that this framework can be directly deployed in scenarios with constrained training resources or when working with limited collections of high-value instruction data.

6. Integration with Supervised Fine-Tuning Pipelines

Q-Tuning provides a practical, scalable blueprint for practitioners seeking to optimize data curation for LLM SFT. The approach requires only access to model PPL and entropy statistics and can be integrated into existing SFT loops with minimal modification. The dynamic nature of both sample and token selection allows the method to adapt to model improvement over the course of training.

As the first dynamic pruning method reported to consistently outperform full-data training, Q-Tuning addresses the growing importance of data efficiency as SFT rivals pretraining in scale, and serves as a reference framework for future work on unified data selection and pruning in LLM fine-tuning.