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

This paper introduces the ODA-Fin datasets and demonstrates that prioritizing high-quality Chain-of-Thought distillation and difficulty-aware sampling in post-training significantly enhances the performance of financial Large Language Models, enabling an 8B-parameter model to surpass state-of-the-art open-source financial LLMs across diverse benchmarks.

The Gist

Imagine you are trying to teach a brilliant but inexperienced apprentice how to become a master financial analyst. You have two choices:

1. The "Model-Centric" approach: Buy the most expensive, massive textbook library (a huge AI model) and hope that simply reading more pages makes them smarter.
2. The "Data-Centric" approach (This Paper): Take a standard, capable apprentice and feed them a curated, high-quality training manual that teaches them exactly how to think, not just what to memorize.

This paper, titled "Unlocking Data Value in Finance," argues that in the complex world of finance, how you prepare the training data is far more important than how big the AI model is.

Here is the story of how the researchers did it, broken down into simple steps:

1. The Problem: The "Noisy Library"

Financial data is messy. Imagine a library where 80% of the books are just simple Q&A flashcards ("What is a stock?"), and the rest are messy, unverified notes with no explanation of how the answer was found.

• The Issue: If you train an AI on this "raw" data, it learns to guess or memorize facts. But finance requires reasoning (e.g., "If interest rates go up, how does this specific bond price change?").
• The Risk: In finance, a small mistake isn't just a typo; it can cost millions of dollars. The AI cannot afford to "hallucinate" (make things up).

2. The Solution: The "Master Chef" Recipe

The researchers didn't build a new AI architecture. Instead, they acted like Master Chefs preparing a gourmet meal from raw ingredients. They created two special datasets:

Phase A: The "SFT" Dataset (The Cookbook)

• Name: ODA-Fin-SFT-318k
• The Process: They took thousands of raw financial questions and ran them through a "distillation" process.
  • Step 1: Clean Up. They removed duplicate questions (like throwing away 100 copies of the same recipe).
  • Step 2: Add the "Why". For questions that just had an answer, they used a super-smart AI to write out the step-by-step reasoning (Chain-of-Thought). It's like turning a recipe that just says "Make a cake" into one that says "Mix eggs, then fold in flour, then bake at 350°F."
  • Step 3: Taste Test. They used a strict "verifier" to check every single step. If the math was wrong or the logic was shaky, the recipe was thrown out.
• The Result: A clean, 318,000-item cookbook where every answer comes with a verified, logical explanation.

Phase B: The "RL" Dataset (The Hard Exam)

• Name: ODA-Fin-RL-12k
• The Process: Once the apprentice learned the basics from the cookbook, they needed to be challenged.
  • They selected only the hardest questions that the AI got wrong more than 50% of the time.
  • Crucial Rule: These hard questions had to be verifiable. You can't ask the AI to "write a poem about the economy" because it's hard to grade. You ask, "What is the exact profit margin?" because there is a right or wrong answer.
• The Result: A 12,000-item "Final Exam" designed to push the AI to think harder without getting confused by ambiguous questions.

3. The Training: From Apprentice to Expert

They took a standard, mid-sized AI model (Qwen3-8B) and trained it using this new data:

1. SFT (Supervised Fine-Tuning): The model studied the "Cookbook" (318k high-quality reasoning steps). It learned how to think logically.
2. RL (Reinforcement Learning): The model took the "Final Exam" (12k hard, verifiable questions). Every time it got a step right, it got a reward. Every time it took a shortcut, it got a penalty.

4. The Results: Small Model, Big Brain

The results were surprising. The researchers trained a small model (8 billion parameters) using this high-quality data.

• The Comparison: They compared it to massive, expensive models and other specialized financial AIs.
• The Outcome: Their small model beat almost everyone.
  • It solved complex math and table problems better than models four times its size.
  • It understood financial news sentiment better than models trained on proprietary (secret) data.
  • Key Insight: A small model with a perfectly curated diet outperformed a giant model with a junk food diet.

5. The Big Lesson: Quality Over Quantity

The paper concludes with a powerful metaphor for the future of AI:

• Old Way: "Let's make the model bigger and bigger!" (Model-Centric)
• New Way: "Let's make the data cleaner, harder, and more logical!" (Data-Centric)

They found that dumping more raw data on a smart model actually makes it worse (it gets confused by noise). But giving it less data that is perfectly verified and explained makes it a genius.

In short: In the high-stakes world of finance, you don't need a bigger brain; you need a better teacher. This paper proves that if you teach the AI how to reason with high-quality, verified examples, even a modest AI can become a financial wizard.

Technical Summary

1. Problem Statement

While Large Language Models (LLMs) excel in general domains, their deployment in finance faces significant hurdles:

• Domain Specificity: High density of specialized terminology and strict factual requirements.
• Reasoning Rigor: Financial tasks demand precise numerical reasoning and multi-step logical inference.
• Error Intolerance: Low tolerance for factual hallucinations, which can lead to severe economic consequences.
• Data Quality Gap: Existing open-source financial datasets often suffer from redundancy, lack of explicit reasoning steps (Chain-of-Thought), and inconsistent quality.
• Paradigm Limitation: The field has historically focused on "Model-Centric" approaches (scaling parameters or architectural changes) rather than addressing the root cause: the quality and structure of post-training data.

2. Methodology

The authors propose a Data-Centric approach, arguing that in specialized verticals, performance is determined by the quality, difficulty, and verifiability of the training data. They introduce a two-stage training pipeline:

A. Data Engineering & Construction

The team constructed two complementary datasets from over 25 open-source repositories (totaling ~697k raw samples):

1. ODA-Fin-SFT-318k (Supervised Fine-Tuning):
   • Goal: Establish a robust foundation for instruction following and domain knowledge.
   • Pipeline:
     • Semantic Deduplication: Used embedding models to remove redundant samples.
     • Reasoning Synthesis: Applied multi-stage distillation. For raw Q&A pairs, a powerful reasoning model (Qwen3-235B) generated step-by-step Chain-of-Thought (CoT) traces.
     • Length-Adaptive Verification: Used a hybrid verification strategy. Short answers were verified by a rule-based/reward model (CompassVerifier-7B), while long-form analysis was verified by a stronger LLM judge.
     • Filtering: Removed samples where the verifier failed or where the model could not solve the problem, ensuring only high-quality, correct reasoning traces remained.
2. ODA-Fin-RL-12k (Reinforcement Learning):
   • Goal: Push the model beyond the SFT plateau by focusing on "hard-but-verifiable" tasks.
   • Selection Criteria:
     • Difficulty: Samples where the SFT model failed >50% of the time were selected.
     • Verifiability: To ensure precise reward signals, samples were filtered to have concise final answers (token length < 16). This allows for reliable automated verification without the noise of long-form generation.

B. Training Pipeline

• SFT Stage: Trained on the 318k distilled CoT dataset using standard full-parameter fine-tuning.
• RL Stage: Used Group Relative Policy Optimization (GRPO).
  • Reward Function: A multiplicative reward combining Format Reward (correct use of <thought> tags) and Correctness Reward (verified by CompassVerifier-7B).
  • Strategy: The model is penalized for incorrect answers but given partial credit for correct formatting to avoid sparse rewards during early training.

3. Key Contributions

1. ODA-Fin-SFT-318k: A curated dataset of 318k high-quality CoT samples derived from open-source data via rigorous distillation and verification, replacing raw data with verified reasoning traces.
2. ODA-Fin-RL-12k: A specialized RL dataset of 12k samples selected for high difficulty and verifiability (short answers), optimizing the trade-off between reward precision and task diversity.
3. Empirical Validation of Data-Centric AI: The study demonstrates that data quality supersedes model scaling in vertical domains. They achieved SOTA performance using an 8B model trained on distilled data, outperforming larger general models and specialized financial models.
4. Ablation Insights:
   • Raw Data Harm: Training capable models (e.g., Qwen3-8B) on raw, unfiltered data degrades performance compared to the base model.
   • Full Distillation is Key: Replacing raw data entirely with distilled CoT yields consistent gains; mixing partial CoT with raw data leads to unstable performance.
   • Negative Transfer: Adding general math or non-financial CoT data to the financial corpus actually reduced performance, highlighting the need for domain-aligned data genealogy.

4. Experimental Results

The models were evaluated on 9 benchmarks across General Financial Understanding, Sentiment Analysis, and Numerical Reasoning.

• Model Performance: ODA-Fin-RL-8B achieved an average score of 74.6%, surpassing all open-source financial LLMs of comparable size (e.g., Fin-R1-7B, DianJin-R1-7B) and even outperforming the much larger Qwen3-32B (74.7%) in several metrics.
• Numerical Reasoning: The model showed dominant performance on TaTQA (89.3%), surpassing the Qwen3-32B baseline by 4.2 points and the DianJin-R1-7B by 7.3 points.
• General Understanding: Outperformed Qwen3-8B by 3.1 points on average and achieved the best Finova score (54.6%) among 8B models.
• Ablation Findings:
  • Starting RL from a strong SFT initialization (ODA-Fin-SFT-8B) was crucial; starting from the base model caused performance degradation.
  • A 16-token answer constraint for RL data filtering provided the optimal balance between verification precision and task diversity, outperforming both 1-token (too narrow) and 128-token (too noisy) constraints.

5. Significance and Implications

• Paradigm Shift: The paper provides strong empirical evidence that in high-stakes verticals like finance, data engineering is more critical than algorithmic novelty or parameter scaling.
• Cost-Effectiveness: It proves that high-performance financial AI can be built using entirely open-source resources without reliance on proprietary data, provided the data undergoes rigorous distillation and verification.
• Methodological Blueprint: The "Hard-but-Verifiable" RL strategy offers a blueprint for training reasoning models in other domains where factual accuracy is paramount.
• Open Release: The authors release the ODA-Fin-SFT-318k, ODA-Fin-RL-12k, and the trained ODA-Fin-8B models to foster reproducible, data-centric research in financial AI.

In conclusion, the paper demonstrates that by transforming raw financial data into high-quality, verified reasoning traces and applying difficulty-aware RL, it is possible to create specialized financial LLMs that outperform significantly larger general-purpose models.