Source-linked AI summary
Unlocking Data Value in Finance: A Study on Distillation and Difficulty-Aware Training
Chuxue Cao, Honglin Lin, Zhanping Zhong, Xin Gao, Mengzhang Cai, Conghui He, Sirui Han, Lijun Wu
TL;DR
Financial LLM deployment requires precise reasoning over specialized terminology while tolerating few factual errors, but the field lacks a systematic account of high-quality financial post-training data. The paper conducts a controlled data-centric study, using verified CoT distillation for SFT and hard-but-verifiable sampling for RL. Across nine benchmarks, ODA-Fin-RL-8B achieves SOTA results among comparable open-source financial LLMs, and the authors release the datasets and trained models.
Problem
Financial tasks demand specialized terminology, numerical reasoning, and reliable factual outputs, while systematic understanding of high-quality financial adaptation data remains limited.
Method
The study constructs ODA-Fin-SFT-318k through multi-stage CoT distillation and verification, and ODA-Fin-RL-12k through difficulty- and verifiability-aware selection using standard SFT and RL pipelines.
Results
ODA-Fin-RL-8B achieves SOTA results across nine financial benchmarks, with particularly strong improvements in numerical reasoning tasks.
Takeaways & Limitations
The findings support data quality, difficulty, and verifiability as key determinants of performance ceilings in vertical financial domains.
Abstract
from arXiv · showhide
Large Language Models (LLMs) have demonstrated strong general capabilities, yet their deployment in finance remains challenging due to dense domain-specific terminology, stringent numerical reasoning requirements, and low tolerance for factual errors. We conduct a controlled empirical study showing that in specialized vertical domains, performance is largely determined by the quality and difficulty/verifiability profile of post-training data. We introduce \textbf{ODA-Fin-SFT-318k}, constructed via multi-stage distillation and verification to produce high-quality Chain-of-Thought supervision, and \textbf{ODA-Fin-RL-12k}, curated for hard-but-verifiable tasks that balance reward precision and task diversity. Using standard SFT and RL pipelines, we show that high-quality CoT distillation establishes a robust foundation during SFT, while difficulty- and verifiability-aware sampling improves RL generalization. Evaluated on nine benchmarks spanning general financial tasks, sentiment analysis, and numerical reasoning, our ODA-Fin-RL-8B consistently surpasses open-source state-of-the-art (SOTA) financial LLMs of comparable size. We release our ODA-Fin-SFT-318k and ODA-Fin-RL-12k datasets, along with trained models to advance data-centric financial AI research.
1 Introduction
Finance remains difficult for LLMs because it combines dense terminology, demanding numerical reasoning, and low tolerance for factual errors. This study examines how post-training data quality, difficulty, and verifiability shape financial-model performance.
- Financial deployment is challenging because tasks require domain-specific terminology, precise numerical reasoning, and reliable factual outputs.
- The field lacks a systematic account of what constitutes high-quality data for financial domain adaptation.
- The study argues that SFT benefits from pure data and high-quality reasoning traces, whereas RL benefits from hard-but-verifiable samples.
- ODA-Fin-SFT-318k uses multi-stage distillation and verification to provide high-quality Chain-of-Thought supervision at scale.
- ODA-Fin-RL-12k is curated for difficult, verifiable tasks, and the study reports that suitable answer-length constraints support stable RL gains.
- The released datasets and trained models are evaluated across financial understanding, sentiment analysis, and numerical reasoning benchmarks.
2 Data Engineering
The data-engineering pipeline converts diverse open-source financial data into curated SFT and RL corpora. It removes redundancy, adds and verifies reasoning traces, filters lengths, and selects difficult but reliably verifiable RL samples.
- 2.1 Data Source: The raw collection aggregates financial datasets from HuggingFace and GitHub across sentiment analysis, reasoning, forecasting, and quantitative trading.The study reports 697,034 samples from over 25 repositories in the Raw Pool, while another source-diversity passage reports 946k raw samples.
- 2.1 Data Source: Financial QA comprises roughly 80% of the collected data, while specialized tasks such as forecasting, trading, and risk analysis are negligible.
- 2.2 Data Distillation and Filtering: The SFT pipeline addresses redundancy and missing reasoning through semantic deduplication, Chain-of-Thought synthesis, verification, and length filtering.
- 2.2 Data Distillation and Filtering: Generated reasoning is verified with different judges for short-form and long-form answers, and samples that fail verification are excluded.
- 2.2 Data Distillation and Filtering: Samples exceeding 16,384 tokens are removed before processed data are merged into ODA-Fin-SFT-318k.
- 2.2 Data Distillation and Filtering: For RL, samples with failure rates greater than 50% are retained, then restricted to final answers shorter than 16 tokens for reliable online verification.
3 Model Training
Training uses standard SFT followed by RL, with verified reasoning traces providing the SFT foundation and lightweight verifier-based rewards guiding RL. The reward combines response-format compliance with extracted-answer correctness.
- The training pipeline applies SFT to distilled reasoning data and then RL to difficulty-aware data.
- SFT replaces raw answers lacking reasoning with verified distilled Chain-of-Thought traces and merges them with existing high-quality CoT datasets.
- The RL stage uses CompassVerifier-7B for low-latency online rewards and restricts examples to concise ground-truth answers.
- The format reward scores whether the response contains correctly ordered <think> tags, using start, end, and pair indicators.
- Correctness is assessed after hierarchical answer extraction, with rule-based matching or verifier judgments against the ground truth.
- The total reward multiplies format and outcome scores, granting partial credit when reasoning format is correct despite an incorrect answer.
4 Experiments
Across nine benchmarks, ODA-Fin-RL-8B achieves strong overall and task-specific performance, with gains over comparable financial models and larger general-purpose baselines. Ablations show that complete distilled CoT supervision and carefully filtered RL data are central to these results.
- Overall Performance and Versatility: 74.6% average performance makes ODA-Fin-RL-8B the highest-scoring open-source model at the 8B scale and nearly matches Qwen3-32B at 74.7%.It also surpasses specialized financial LLMs of comparable size.
- Numerical Reasoning Capabilities: 89.3% on TaTQA, 80.4% on ConvFinQA, and 73.3% on FinQA demonstrate dominant numerical-reasoning performance.On TaTQA, ODA-Fin-RL-8B exceeds Qwen3-8B by 2.2 points, Dianjin-R1-7B by 7.3 points, and Qwen3-32B by 4.2 points.
- General Financial Understanding: 54.6% on Finova is the best result among 8B-scale models, while robust scores across general-understanding benchmarks preserve domain knowledge.Dianjin-R1-7B leads on FinEval and FinanceIQ, and Qwen3-32B reaches 80.8% on FinanceIQ.
- Sentiment Analysis: 83.4% on FPB and 78.5% on Headlines outperform Qwen3-32B on those tasks, while specialized models retain higher scores on selected sentiment benchmarks.The results indicate balanced performance across sentiment tasks but leave room to improve task-specific data mixtures.
- Impact of Reinforcement Learning: 72.1% to 74.6% average performance shows consistent gains from RL over SFT, including +6.8 points on Finova and +3.5 points on FinQA.The comparison attributes these gains to improved instruction-following and reasoning capabilities after RL.
- SFT Data Composition: Complete distilled CoT supervision achieves the best SFT performance across both architectures, reaching 71.0% for Qwen2.5 and 72.1% for Qwen3.Mixed or table-calculation augmentation lowers performance to 67.6% and 66.0%, respectively.
- RL Data and Reward Designs: 74.6% average accuracy is achieved with the model-based verifier and 16-token constraint, outperforming the 73.3% rule-based configuration.A 128-token constraint increases diversity but reduces performance to 73.1%, illustrating a precision–diversity trade-off.
5 Discussion & Insights
The discussion argues that financial LLM performance depends primarily on rigorous data quality and training-strategy design rather than scale alone. It highlights gains from distilled reasoning data and targeted reinforcement learning, alongside risks from indiscriminate out-of-domain augmentation.
- Data quality and training strategies are presented as more important than model-centric scaling for financial LLM adaptation.The authors frame this as a shift toward data-centric AI in finance.
- Distilled, high-quality Chain-of-Thought data outperformed mixtures containing raw or general data in the reported experiments.The discussion attributes this pattern to rigorous processing rather than dataset volume alone.
- 89.3% TaTQA performance was achieved by ODA-Fin-RL-8B using explicit CoT generation and difficulty-oriented RL on samples with ≥50% failure rate.The reward design used a 16-token verifier constraint to balance precision and diversity.
- Mixing DeepMath and Ring-Light with the financial corpus harmed performance, indicating negative transfer from domain distribution shifts.The discussion links this to finance-specific conventions and implicit context distinct from pure mathematical logic.
6 Conclusion
The conclusion reports that data quality, difficulty, and verifiability determine performance ceilings in vertical financial domains. Controlled SFT and RL experiments support high-quality CoT distillation and difficulty-aware sampling as the central ingredients of the approach.
- ODA-Fin-RL-8B achieves SOTA results across nine benchmarks, with particularly strong improvements in numerical reasoning tasks.The study evaluates general financial tasks, sentiment analysis, and numerical reasoning.
- High-quality CoT distillation establishes a robust instruction-following foundation, while difficulty- and verifiability-aware RL improves generalization.The RL strategy balances reward precision and task diversity.
- The authors release ODA-Fin-SFT-318k, ODA-Fin-RL-12k, and trained models for data-centric financial AI research.
A Dataset Details
The dataset details describe an overview of financial instruction datasets and identify the data composition and task-category distribution of ODA-Fin-RL-12k. Table 4 includes dataset sources, sizes, languages, CoT availability, descriptions, and download counts.
- Table 4 overviews financial instruction datasets by year, size, language, CoT availability, description, download counts, and link.Download counts are reported from Hugging Face as of February 2026.
- Figure 5 presents the data-source composition and task-category distribution for ODA-Fin-RL-12k.
B Prompts
The prompt appendix presents templates for generating Chain-of-Thought responses and evaluating long-form financial answers. The evaluation criteria emphasize financial concepts, numerical accuracy, completeness, logical reasoning, and acceptable semantic variations.
- CoT generation: Listing 1 uses Qwen3-235-A22B-Thinking to generate step-by-step financial reasoning with a boxed final answer.
- Answer verification: Listing 2 evaluates whether a student answer is semantically consistent with a reference answer in finance and accounting.
- Evaluation criteria: The evaluator checks core financial concepts, numerical accuracy, and technical completeness, including key financial metrics and indicators.
- Evaluation criteria: The rubric additionally requires sound financial analysis and reasoning while allowing equivalent wording, terminology, formats, and minor rounding differences.
- Answer verification: The verification prompt returns Yes for semantically consistent answers and No for inconsistent answers or material errors.