Rajveer-code commited on
Commit
c1d7199
·
verified ·
1 Parent(s): 0292e40

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +180 -2
README.md CHANGED
@@ -31,8 +31,186 @@ configs:
31
  data_files:
32
  - split: test
33
  path: data/test-*
34
- - split: validation
35
- path: data/validation-*
36
  - split: dev
37
  path: data/dev-*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31
  data_files:
32
  - split: test
33
  path: data/test-*
 
 
34
  - split: dev
35
  path: data/dev-*
36
+ license: cc-by-4.0
37
+ task_categories:
38
+ - question-answering
39
+ - text-classification
40
+ language:
41
+ - en
42
+ tags:
43
+ - finance
44
+ - legal
45
+ - regulatory
46
+ - india
47
+ - benchmark
48
+ - llm-evaluation
49
+ - sebi
50
+ - rbi
51
+ pretty_name: IndiaFinBench
52
+ size_categories:
53
+ - n<1K
54
  ---
55
+
56
+ # IndiaFinBench
57
+
58
+ **An Evaluation Benchmark for Large Language Model Performance on Indian Financial Regulatory Text**
59
+
60
+ > Rajveer Singh Pall — Gyan Ganga Institute of Technology and Sciences, Jabalpur, India
61
+
62
+ [![GitHub](https://img.shields.io/badge/GitHub-IndiaFinBench-blue)](https://github.com/rajveerpall/IndiaFinBench)
63
+ [![Paper](https://img.shields.io/badge/arXiv-IndiaFinBench-red)](https://arxiv.org/abs/XXXX.XXXXX)
64
+ [![License: CC BY 4.0](https://img.shields.io/badge/License-CC%20BY%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by/4.0/)
65
+
66
+ ---
67
+
68
+ ## Dataset Summary
69
+
70
+ IndiaFinBench is, to our knowledge, the first publicly available evaluation benchmark for assessing large language model (LLM) performance on **Indian financial regulatory text**. Existing financial NLP benchmarks draw exclusively from Western corpora — SEC filings, US earnings reports, and English-language financial news — leaving a significant gap in coverage of non-Western regulatory frameworks.
71
+
72
+ IndiaFinBench addresses this gap with **406 expert-annotated question-answer pairs** drawn from **192 regulatory documents** sourced directly from the Securities and Exchange Board of India (SEBI) and the Reserve Bank of India (RBI), spanning documents from 1992 to 2026.
73
+
74
+ The benchmark covers four task types that probe distinct reasoning capabilities required for Indian regulatory text:
75
+
76
+ | Task Type | Code | Items | Description |
77
+ |-----------|------|-------|-------------|
78
+ | Regulatory Interpretation | REG | 174 | Identify correct rules, thresholds, or scope from regulatory passages |
79
+ | Numerical Reasoning | NUM | 92 | Perform arithmetic over figures embedded in regulatory text |
80
+ | Contradiction Detection | CON | 62 | Determine whether two regulatory passages contradict each other |
81
+ | Temporal Reasoning | TMP | 78 | Order regulatory events, identify which circular was operative at a given time |
82
+ | **Total** | | **406** | |
83
+
84
+ ---
85
+
86
+ ## Key Results
87
+
88
+ Twelve models evaluated under zero-shot conditions on the full 406-item benchmark:
89
+
90
+ | Model | REG | NUM | CON | TMP | Overall |
91
+ |-------|-----|-----|-----|-----|---------|
92
+ | Gemini 2.5 Flash | 93.1% | 84.8% | 88.7% | 88.5% | **89.7%** |
93
+ | Qwen3-32B | 85.1% | 77.2% | 90.3% | 92.3% | 85.5% |
94
+ | LLaMA-3.3-70B | 86.2% | 75.0% | 95.2% | 79.5% | 83.7% |
95
+ | Llama 4 Scout 17B | 86.2% | 66.3% | 98.4% | 84.6% | 83.3% |
96
+ | Kimi K2 | 89.1% | 65.2% | 91.9% | 75.6% | 81.5% |
97
+ | LLaMA-3-8B | 79.9% | 64.1% | 93.5% | 78.2% | 78.1% |
98
+ | GPT-OSS 120B | 79.9% | 59.8% | 95.2% | 76.9% | 77.1% |
99
+ | GPT-OSS 20B | 79.9% | 58.7% | 95.2% | 76.9% | 76.8% |
100
+ | Gemini 2.5 Pro | 89.7% | 48.9% | 93.5% | 64.1% | 76.1% |
101
+ | Mistral-7B | 79.9% | 66.3% | 80.6% | 74.4% | 75.9% |
102
+ | DeepSeek R1 70B | 72.4% | 69.6% | 96.8% | 70.5% | 75.1% |
103
+ | Gemma 4 E4B | 83.9% | 50.0% | 72.6% | 62.8% | 70.4% |
104
+ | **Human Baseline (non-specialist)** | 55.6% | 44.4% | 83.3% | 66.7% | **60.0%** |
105
+
106
+ All models substantially outperform the non-specialist human baseline. Numerical reasoning is the most discriminative task (35.9 percentage-point spread across models).
107
+
108
+ ---
109
+
110
+ ## Dataset Details
111
+
112
+ ### Source Documents
113
+
114
+ | Source | Documents | Types |
115
+ |--------|-----------|-------|
116
+ | SEBI (sebi.gov.in) | 92 | Circulars, master circulars, regulations, orders |
117
+ | RBI (rbi.org.in) | 100 | Circulars, monetary policy statements, master directions |
118
+ | **Total** | **192** | |
119
+
120
+ ### Difficulty Distribution
121
+
122
+ | Difficulty | Items | Description |
123
+ |------------|-------|-------------|
124
+ | Easy | 160 (39.4%) | Single-step extraction from context |
125
+ | Medium | 182 (44.8%) | Multi-clause reasoning or calculation |
126
+ | Hard | 64 (15.8%) | Multi-instrument tracking or complex arithmetic |
127
+
128
+ ### Splits
129
+
130
+ The dataset is split into **test** (324 items, 79.8%) and **dev** (82 items, 20.2%).
131
+
132
+ | Split | Items |
133
+ |-------|-------|
134
+ | test | 324 |
135
+ | dev | 82 |
136
+ | **Total** | **406** |
137
+
138
+ ---
139
+
140
+ ## Data Fields
141
+
142
+ | Field | Type | Description |
143
+ |-------|------|-------------|
144
+ | `id` | string | Unique item identifier (e.g., `REG_001`, `NUM_042`) |
145
+ | `task_type` | string | One of: `regulatory_interpretation`, `numerical_reasoning`, `contradiction_detection`, `temporal_reasoning` |
146
+ | `difficulty` | string | One of: `easy`, `medium`, `hard` |
147
+ | `source` | string | Regulatory body: `SEBI` or `RBI` |
148
+ | `context` | string | Regulatory passage(s) provided to the model (80–500 words). For contradiction detection items, contains Passage A and Passage B separated by a delimiter |
149
+ | `question` | string | The question to be answered from the context |
150
+ | `reference_answer` | string | Gold-standard reference answer |
151
+ | `source_document` | string | Filename of the source regulatory document |
152
+
153
+ ---
154
+
155
+ ## Annotation and Validation
156
+
157
+ All 406 QA pairs were authored by a domain expert in Indian financial regulation. Every item was individually reviewed to ensure:
158
+ - The answer is unambiguously derivable from the provided context
159
+ - The question has exactly one correct answer
160
+ - The context is sufficient without external knowledge
161
+
162
+ **Model-based secondary validation** (LLaMA-3.3-70B, 150-item subset): 90.7% agreement, κ = 0.918 on contradiction detection.
163
+
164
+ **Human inter-annotator agreement** (second human annotator, 60-item sample): 76.7% overall agreement, κ = 0.611 for contradiction detection (substantial agreement per Landis & Koch 1977).
165
+
166
+ ---
167
+
168
+ ## Evaluation Protocol
169
+
170
+ Models are evaluated under **zero-shot, context-only** conditions. The scoring pipeline applies four stages in sequence:
171
+
172
+ 1. **Exact match** after case-normalisation and punctuation stripping
173
+ 2. **Fuzzy token match** using RapidFuzz `token_set_ratio ≥ 0.72`
174
+ 3. **Numerical extraction match** for items where extracted number sets agree
175
+ 4. **Yes/No match** for contradiction detection (leading word comparison)
176
+
177
+ Full evaluation code and all model predictions are available at: [https://github.com/rajveerpall/IndiaFinBench](https://github.com/rajveerpall/IndiaFinBench)
178
+
179
+ ---
180
+
181
+ ## Limitations
182
+
183
+ - All evaluation is zero-shot; few-shot or chain-of-thought prompting may improve performance
184
+ - The benchmark does not currently cover Hindi–English code-switched regulatory text
185
+ - Coverage is limited to SEBI and RBI; extension to IRDAI, PFRDA, and commodity regulation is planned
186
+ - The benchmark evaluates short extractive responses, not longer-form regulatory reasoning or document summarisation
187
+ - The dataset is a snapshot of documents as of early 2026; regulatory frameworks evolve continuously
188
+
189
+ ---
190
+
191
+ ## Citation
192
+
193
+ If you use IndiaFinBench in your research, please cite:
194
+
195
+ ```bibtex
196
+ @article{pall2025indiafinbench,
197
+ title={IndiaFinBench: An Evaluation Benchmark for Large Language Model Performance on Indian Financial Regulatory Text},
198
+ author={Pall, Rajveer Singh},
199
+ journal={arXiv preprint},
200
+ year={2025},
201
+ url={https://github.com/rajveerpall/IndiaFinBench}
202
+ }
203
+ ```
204
+
205
+ ---
206
+
207
+ ## License
208
+
209
+ This dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). All source documents are publicly available from sebi.gov.in and rbi.org.in and carry no copyright restrictions on research use.
210
+
211
+ ---
212
+
213
+ ## Contact
214
+
215
+ Rajveer Singh Pall — rajveer.singhpall.cb23@ggits.net
216
+ Gyan Ganga Institute of Technology and Sciences, Jabalpur, India