Diagnosing Capability Gaps in Fine-Tuning Data

1. Problem Statement

Fine-tuning Large Language Models (LLMs) for domain-specific tasks (e.g., medical QA, legal summarization, code generation) requires training datasets that comprehensively cover the specific capabilities needed. However, practitioners currently lack a principled way to determine if a dataset supports a high-level goal before committing to an expensive fine-tuning run.

Existing methods rely on aggregate metrics (perplexity, BLEU) or post-hoc performance analysis, which fail to:

• Localize specific capability gaps (e.g., a model might summarize well generally but fail on monetary details in legal bills).
• Identify undersupported sub-goals until after production failures occur.
• Provide actionable signals for dataset curation or synthetic data generation.

2. Methodology: The GOALCOVER Framework

The authors introduce GOALCOVER, a four-phase framework designed to proactively detect capability gaps through interactive goal decomposition and automated coverage assessment.

Phase 1: Interactive Goal Clarification

Instead of accepting a vague high-level goal, the framework guides the practitioner through a structured elicitation protocol:

• Decomposition: The system breaks a high-level goal $G$ into a set of atomic, independently evaluable subgoals $S = \{s_1, ..., s_k\}$.
• Constraints: The decomposition enforces atomicity (each subgoal is a single capability) and disjointness (subgoals do not overlap).
• Process: An iterative loop of clarifying questions and practitioner responses ensures the subgoals map to specific model capabilities (e.g., "Cardiology expertise" vs. "General clinical reasoning").

Phase 2: Coverage Assessment

• Alignment Scoring: An LLM-based evaluator ($f_{eval}$) scores every training sample $(x_i, y_i)$ against every subgoal $s$.
• Rubric: The evaluator uses an anchored-rubric prompt (0.0 to 1.0 scale) to produce a quantitative alignment score and a structured natural language explanation ($r_i$) justifying the score.
• Aggregation: The system calculates the average alignment score $\bar{A}(s)$ for each subgoal across the dataset.

Phase 3: Gap Analysis

• Identification: Subgoals with low average scores ($\bar{A}(s) < \tau$) are flagged as capability gaps.
• Synthesis: The system aggregates the structured explanations from low-scoring samples to identify recurring missing capabilities (e.g., "lack of drug interaction warnings").

Phase 4: Recommendation & Remediation

• Actionable Feedback: The system provides specific curation suggestions (e.g., "add cardiology-specific examples").
• Synthetic Data Generation: It employs a Generator-Discriminator pipeline to create goal-conditioned synthetic samples. The generator creates data targeting the identified gaps, and the discriminator filters them based on alignment with the specific subgoal.

3. Key Contributions

1. Structured Goal Decomposition: A principled methodology to break complex goals into atomic, disjoint subgoals via a fixed elicitation protocol.
2. LLM-Based Alignment Assessment: A scoring framework that produces both quantitative scores and qualitative explanations for why a sample fails a specific capability.
3. Adversarial Validation: A controlled corruption protocol (data ablation) that validates the detection mechanism by selectively removing content and verifying that only the targeted subgoal's score drops significantly.
4. Cross-Domain Empirical Validation: Demonstrated effectiveness across Medical QA, Legal Summarization, and Code Generation.
5. Downstream Utility: Proven improvement in Reinforcement Fine-Tuning (RFT) outcomes using filtered and synthetic data.
6. Practical Pipeline: An end-to-end workflow providing specific, actionable recommendations for closing data gaps.

4. Experimental Results

Validation 1: Controlled Corruption (Detection Validity)

The authors simulated capability gaps by removing specific content types from datasets (e.g., removing "cardiac" terms from medical QA).

• Target vs. Non-Target Separation: When content supporting a specific subgoal was removed, that subgoal's alignment score degraded by an average of 25.6%, whereas non-target subgoals degraded by only 2.1%.
• Statistical Significance: The separation is statistically significant ($p < 0.001$) with a large effect size (Cohen's $d = 1.24$).
• Domain Specifics:
  • Legal: Removing financial content caused a 30.75% drop in "Monetary preservation" scores.
  • Medical: Removing cardiology content caused a 26.28% drop in "Cardiology expertise" scores.
  • Code: Removing ML library references caused a 15.53% drop in "ML libraries" scores.

Validation 2: Downstream Utility (RFT Experiments)

The framework was tested on a Financial Summarization task using Qwen-3-14B and GRPO (Group Relative Policy Optimization).

• Baseline: Unfiltered data achieved an LLM-judge reward of 3.77.
• GOALCOVER Filtered: Using only data aligned with subgoals improved the score to 4.12 (despite using fewer samples).
• Synthetic Data: Goal-conditioned synthetic data alone achieved 4.02.
• Best Result: Combining filtered real data with goal-conditioned synthetic data achieved the highest score of 4.20.
• Subgoal Consistency: The improvements were consistent across 10 different input distribution subgoals (e.g., long documents, noisy inputs, specific policy types), with no regressions observed.

5. Significance and Impact

• Proactive vs. Reactive: GOALCOVER shifts dataset evaluation from a reactive post-training analysis to a proactive pre-training diagnostic, saving significant computational resources by preventing failed fine-tuning runs.
• Granular Visibility: It moves beyond aggregate metrics to provide subgoal-level visibility, allowing practitioners to pinpoint exactly which capabilities are missing (e.g., "The model knows law but not money").
• Closed-Loop Optimization: By linking gap detection directly to synthetic data generation, the framework offers a complete solution for not just finding, but also fixing data deficiencies.
• Reliability: The controlled corruption experiments provide strong evidence that the LLM-based evaluators are detecting genuine capability dependencies rather than random noise.

In conclusion, GOALCOVER provides a robust, interactive framework for ensuring that fine-tuning datasets are aligned with specific, high-stakes domain goals, significantly improving model reliability and safety before deployment.