Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

It is shown that high-quality CoT distillation establishes a robust foundation during SFT, while difficulty- and verifiability-aware sampling improves RL generalization, and the ODA-Fin-RL-8B consistently surpasses open-source state-of-the-art (SOTA) financial LLMs of comparable size.

Abstract

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.

Unlocking Data Value in Finance: A Study on Distillation and Difficulty-Aware Training

TL;DR

It is shown that high-quality CoT distillation establishes a robust foundation during SFT, while difficulty- and verifiability-aware sampling improves RL generalization, and the ODA-Fin-RL-8B consistently surpasses open-source state-of-the-art (SOTA) financial LLMs of comparable size.

Abstract

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.
Paper Structure (36 sections, 3 equations, 6 figures, 3 tables)

This paper contains 36 sections, 3 equations, 6 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Data Engineering
  3. Data Source
  4. Source Diversity.
  5. Task Taxonomy.
  6. Data Distillation and Filtering
  7. 1) Semantic Deduplication.
  8. 2) Reasoning Synthesis.
  9. 3) Length-Adaptive Verification.
  10. 4) Token Length Filtering.
  11. Data Selection for RL
  12. Model Training
  13. Supervised Fine-Tuning
  14. Reinforcement Learning
  15. Format Reward.
  16. ...and 21 more sections

Figures (6)

  • Figure 1: Average score across Financial benchmarks. ODA-Fin-RL/SFT-8B demonstrates strong performance relative to thinking models with significantly more parameters.
  • Figure 2: Data construction pipeline of ODA-Fin-SFT-318k and ODA-Fin-RL-12k.
  • Figure 3: Data source and task distribution.
  • Figure 4: Ablation study on answer token length constraints. (a) Data size and source diversity; (b) Token length distribution of reference answers; (c) Average model performance across benchmarks.
  • Figure 5: ODA-Fin-RL-12k data source and task disstribution.
  • ...and 1 more figures