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 ODA-Fin-SFT-318k, constructed via multi-stage distillation and verification to produce high-quality Chain-of-Thought supervision, and 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.
Figures & tables
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.
Model
General
Sentiment Analysis
Numerical
AVG
FinEval
Finova
FinIQ
FOMC
FPB
HL
FinQA
TaTQA
CFQA
Acc/zh
Acc/zh
Acc/zh
W-F1/en
W-F1/en
W-F1/en
Acc/en
Acc/en
Acc/en
General LLMs with Reasoning
GPT-5 mini-high
84.6
67.7
65.9
66.7
77.0
75.4
77.6
92.9
85.7
77.1
Gemini 2.5 Flash-Lite
52.8
51.2
45.7
62.5
79.8
76.0
72.0
90.0
79.5
67.7
Qwen3-4B-Thinking
76.4
54.6
72.3
63.8
70.2
75.0
72.4
87.6
80.6
72.5
Table 1: Main Results: ODA-Fin-RL achieves top three performance across most benchmarks. ‘FinIQ’, ‘HL’ and ‘CFQA’ refer to FinanceIQ, Headlines, and ConvFinQA benchmarks.
Model
Data
General
Sentiment Analysis
Numerical
AVG
FinEval
Finova
FinIQ
FOMC
FPB
HL
FinQA
TaTQA
CFQA
Acc/zh
Acc/zh
Acc/zh
W-F1/en
W-F1/en
W-F1/en
Acc/en
Acc/en
Acc/en
Qwen2.5-7B-Instruct
-
78.9
42.0
65.1
56.4
78.6
73.2
64.6
81.5
72.1
68.0
Qwen2.5-SFT-raw
all 698K raw
75.6
39.6
64.7
70.4
94.1
73.9
64.8
80.5
73.7
70.8
Qwen2.5-SFT-partial
192K partial cot
73.6
43.6
72.2
65.3
74.1
74.6
65.7
82.3
71.7
69.2
Qwen2.5-SFT-all
all 318K cot
75.6
43.2
72.7
64.8
75.3
78.4
68.4
86.6
73.8
71.0
Table 2: SFT Data Ablation Results. ”all cot” (Setting III) achieves the highest average performance. ‘FinIQ’, ‘HL’ and ‘CFQA’ refer to FinanceIQ, Headlines, and ConvFinQA benchmarks.
Model
Ver
#Tok
#Size
General
Sentiment Analysis
Numerical
AVG
FinEval
Finova
FinIQ
FOMC
FPB
HL
FinQA
TaTQA
CFQA
Acc/zh
Acc/zh
Acc/zh
W-F1/en
W-F1/en
W-F1/en
Acc/en
Acc/en
Acc/en
Qwen3-8B
-
-
-
77.8
44.9
72.5
57.5
76.8
76.0
72.2
87.1
78.8
71.5
Fin-Zero-RL
Rule
1
6k
81.5
20.5
72.1
55.7
76.3
75.5
71.3
82.9
80.1
68.4
ODA-Fin-SFT-8B
-
-
-
76.0
47.8
72.1
63.9
75.6
78.2
69.8
87.0
78.3
72.1
Fin-RL-1
Rule
1
6k
79.0
46.7
75.8
63.7
75.1
78.0
72.1
88.9
80.2
73.3
Table 3: Ablation study on RL stage. ‘#Tok’ denotes the maximum final answer’s token length constraint used for filtering RL training data, ‘#Size’ refers the data size. ‘FinIQ’, ‘HL’ and ‘CFQA’ refer to FinanceIQ, Headlines, and ConvFinQA benchmarks.
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.
Dataset
Year
Size
Lang
CoT
Description
DLs
Link
takala/financial_phrasebank [ 19 ]
2013
5K
EN
✗
Financial news sentiment classification
71K
TimKoornstra/financial-tweets-sentiment
2024
38K
EN
✗
Twitter financial sentiment analysis
104
FinGPT/fingpt-sentiment-train
2024
77K
EN
✗
Financial sentiment training data
302
TheFinAI/en-fpb
2024
3k
EN
✗
English financial phrasebank
107
nickmuchi/financial-classification
2022
5k
EN
✗
Financial text classification
129
zeroshot/twitter-financial-news-sentiment
2023
9.5K
EN
✗
Twitter financial news sentiment
4.2K
Table 4: Overview of financial instruction datasets used in this study. Download counts are from Hugging Face as of February 2026.
Figure 5: ODA-Fin-RL-12k data source and task disstribution.