Source-linked AI summary
Data-Centric Post-Training for Financial Reasoning: Mining, Distillation, and Verifiable Learning
Zhirayr Hayrapetyan, Andrei Kalmykov, Denis Kokosinskii, Dmitry Stanishevskii, Dmitry Zmitrovich
TL;DR
Financial reasoning post-training is limited by insufficiently reasoned, self-contained, and verifiable data, while adaptation can erode capabilities already present in starting models. The paper builds complementary corpora with stage-specific classifiers, then studies SFT, self-distillation, merging, and GRPO; across selected FINESSE-Bench comparisons, retention-aware methods outperform ordinary SFT’s regressions.
Problem
Financial reasoning data often lacks explicit reasoning, sufficient context, or reliably verifiable answers, limiting directly usable post-training material.
Method
The paper mines reasoning traces, distills financial instruction data, generates knowledge-graph-guided examples, and applies separate classifiers for relevance, self-containment, and verifiability.
Results
Ordinary SFT reduces FINESSE-Bench accuracy by 3.2–4.0 points, whereas self-distilled SFT improves over starting models by 1.0–2.8 points; merging and GRPO also recover or add performance.
Takeaways & Limitations
Retention-aware adaptation can improve financial reasoning while avoiding the regressions observed after ordinary SFT.
Abstract
from arXiv · showhide
Financial text, textbooks, and question-answer pairs are abundant, but only a small fraction is directly usable for reasoning-focused post-training. Existing QA pairs often lack explicit reasoning, sufficient context, or reliably verifiable answers, while textbooks must first be transformed into synthetic training examples. We present a data-centric pipeline that constructs complementary corpora by mining open-source reasoning traces, distilling financial instruction data, and generating knowledge-graph-guided question-answer pairs from financial educational material. After semantic deduplication, three lightweight sequence classifiers select finance-relevant examples, reject under-specified questions, and identify tasks suitable for reinforcement learning with compact rule-based verifiers. For model adaptation, we study supervised fine-tuning and reinforcement learning, while self-distilled fine-tuning and post-training model merging are used to prevent the loss of financial capabilities already present in the starting model. We evaluate the adapted language models using FINESSE-Bench, reporting aggregate performance and changes relative to their starting checkpoints. Across the selected comparisons, ordinary SFT reduces FINESSE-Bench accuracy by 3.2-4.0 percentage points, whereas self-distilled SFT improves over the corresponding starting models by 1.0-2.8 points. Equal-weight merging recovers 3.0 points over its SFT parent and finishes 0.9 points above the original model; GRPO on hard tasks adds 0.4 points after self-distilled SFT or 3.0 points when applied directly to verifiable tasks. These results show that retention-aware adaptation can improve financial reasoning without the regressions observed after ordinary SFT.
1 Introduction
The paper presents a data-centric pipeline for financial reasoning post-training that treats data construction, verification, and selection as stage-specific problems. It combines complementary corpora with retention-aware adaptation and evaluates the resulting models on FINESSE-Bench.
- Financial deployment requires specialized terminology, numerical accuracy, evidence integration, and auditable conclusions.
- Different post-training stages require distinct data-quality notions rather than one opaque quality score.
- Three classifiers support finance-relevant mining, missing-context rejection, and selection of tasks suitable for compact rule-based rewards.
- The pipeline mines long reasoning traces, distills Finance-Instruct examples, and generates knowledge-graph-guided question–answer pairs from financial educational material.
- 44,652 mined, 4,475 distilled, and 329,828 knowledge-graph-guided synthetic examples form three complementary financial post-training sets.
- Self-distillation and merging mitigate ordinary SFT’s capability loss, with self-distillation finishing 2.8 FINESSE-Bench points above the starting model versus 0.9 for merging.
2 Data Construction
Data construction combines three deduplicated branches designed to balance scale, concept coverage, reasoning diversity, and curation control. The resulting releases and filtered subsets support distinct financial reasoning training needs.
- Three branches combine Glaive reasoning traces, Finance-Instruct data, and GraphGen-generated examples from financial books.
- 645,232 Glaive examples remain after mining, semantic deduplication, response parsing, and exact prompt deduplication.
- 93,570 Glaive examples remain after stringent finance, self-containment, deduplication, prefix, and language filtering.
- 4,475 Finance-Instruct examples remain after language, task-family, finance, self-containment, deduplication, parsing, and single-turn restrictions.
- GraphGen adapts CFA curriculum and exam-preparation books into knowledge graphs and synthesized questions spanning individual facts and compositional concepts.
- The branches trade scale for control: GraphGen provides concept coverage, Glaive diverse long trajectories, and Finance-Instruct jointly curated instruction examples.
3 Auxiliary Data Classifiers
Three lightweight sequence classifiers operationalize separate curation decisions: financial relevance, self-containment, and answer verifiability. Their outputs enable scalable filtering and restrict rule-based reinforcement learning to checkable tasks.
- Each classifier adds a binary head to Qwen/Qwen3-Embedding-0.6B and screens candidates more economically than a generative judge.
- The finance classifier is a mining model whose positive class covers financial domains and whose negative class is non-financial content.
- The missing-context classifier detects questions that cannot be answered from their presented information, using one minus its probability as a self-containment score.
- The verifiability classifier identifies numerical, multiple-choice, or parser-checkable tasks without judging whether a supplied answer is correct.
4 Training Strategies
The training study compares conventional SFT, self-distilled SFT, model merging, and GRPO while treating domain improvement and retention of existing financial behavior as joint objectives.
- Four adaptation strategies are evaluated: conventional SFT, self-distilled SFT, model merging, and reinforcement learning.
- Conventional SFT combines normalized GraphGen, Glaive, and distilled Finance-Instruct conversations with explicit reasoning and final-answer fields where available.
- Self-distilled SFT uses a frozen starting checkpoint as teacher to constrain output-distribution drift while learning new domain facts.
- Equal-weight merging linearly interpolates an SFT checkpoint with its starting checkpoint to recover capabilities lost during fine-tuning.
- GRPO trains either on SFT-hard questions with at least three incorrect sampled responses or on verifiability-selected questions with rule-based rewards.
5 Experiments
The experiments compare SFT, self-distillation, merging, and GRPO on a fixed FINESSE-Bench evaluation set, tracking both aggregate accuracy and changes from starting checkpoints. Results show gains from targeted RL and merging, with experiment chains making adaptation effects directly interpretable.
- Evaluation and experiment design: Experiment chains connect each adapted checkpoint to its starting model and, when applicable, its immediate SFT parent.This structure supports direct comparisons of ordinary SFT, self-distilled SFT, merging, and GRPO across model families and scales.
- Evaluation and experiment design: The study evaluates selected SFT, self-distillation, merging, and RL chains on a fixed 778-question FINESSE-Bench subset.The set spans CFA, CMT, VLigaBench-ru, technical-analysis, derivatives-trading, and FinQA questions, with micro-accuracy reported as the benchmark score.
- RL experiments: 0.39 percentage points: difficulty-selected GRPO raises the self-distilled Gemma-4-12B-IT checkpoint from 79.18 to 79.56 on FINESSE-Bench.The hard-example branch samples questions the checkpoint fails in at least three of four attempts, while component-level effects remain mixed.
- RL experiments: 3.0 percentage points: verifier-compatible RL raises Qwen3-1.7B from 56.0 to 59.0 and Qwen3-4B from 61.0 to 64.0 without preceding domain SFT.This branch uses tasks selected for compact answer verification and applies RL directly to unadapted checkpoints.
- Merging experiments: 2.96 percentage points: equal-weight merging improves the Qwen3.5-4B SFT parent from 73.14 to 76.09 and finishes 0.90 points above the original model.The merge interpolates the SFT checkpoint with its starting checkpoint to recover capabilities lost during fine-tuning.
6 Discussion and Insights
The discussion frames data selection and capability retention as complementary requirements for financial post-training. It also argues that the curated sources provide different forms of reasoning coverage rather than interchangeable examples.
- Stage-specific filtering: Filtering criteria should follow the training stage: finance relevance supports mining, self-containment protects data quality, and verifiability supports defensible RL rewards.The discussion warns that collapsing these signals into one quality score would hide operationally important differences.
- Financial-capability retention: Ordinary SFT reduces accuracy by 3.2 points for Qwen3.5-4B and 4.0 points for Gemma-4-12B-IT, whereas self-distilled SFT improves them by 2.8 and 1.7 points.Equal-weight merging recovers 3.0 points over the Qwen SFT parent and ends 0.9 points above the original model.
- Complementary training sources: GraphGen supplies controlled concept coverage, mined Glaive examples supply diverse long trajectories, and distilled Finance-Instruct supplies a curated instruction-following bridge.These sources address different data limitations and form the supervised-training mixture used in the reported experiments.
7 Conclusion
The paper concludes that explicit, stage-specific curation combined with retention-aware adaptation improves financial reasoning across selected FINESSE-Bench comparisons. Self-distillation, merging, and targeted RL reduce or avoid the regressions associated with ordinary SFT.
- Conclusion: The pipeline makes three explicit curation decisions: finance relevance, self-containment, and compact answer verifiability.These decisions support reasoning mining, Finance-Instruct distillation, and knowledge-graph-guided synthesis from financial educational material.
- Conclusion: Ordinary SFT reduces FINESSE-Bench accuracy by 3.2–4.0 points, whereas self-distilled SFT improves corresponding starting models by 1.0–2.8 points.Equal-weight merging raises the Qwen3.5-4B SFT checkpoint by 3.0 points and finishes 0.9 points above the original model.
- Conclusion: GRPO adds 0.4 points after self-distilled SFT on hard examples and 3.0 points when applied directly to verifiable tasks.Together with explicit data selection and self-distillation, targeted RL improves financial reasoning without the regressions observed under ordinary SFT.