RTI-Bench: A Structured Dataset for Indian RTI Decision Analysis

Dataset Description

RTI-Bench is the first structured dataset of Central Information Commission (CIC) decisions under India's Right to Information Act, 2005. It supports research in legal NLP, civic AI, and AI-assisted access to justice.

Total cases: 1,516 across two sources:

• 1,218 annotated instruction-response pairs (Source A)
• 298 structured CIC PDF decisions spanning 5 commissioners and 3 document format generations, 2023–2026 (Source B)

Overall label coverage: 82.8% of 1,457 primary cases, extracted using a fully reproducible rule-based pipeline — no LLM annotation.

---

Dataset Structure

Source A — Instruction-Response Corpus (hf_annotated)

Derived from jatinmehra/RTI-CASE-DATASET with structured fields added via rule-based extraction.

Field	Type	Description
hf_id	int	Original row index
title	string	RTI case subject line
instruction	string	Background: information sought, PIO response, hearing
response	string	Commission's final direction
public_authority	string	Government department/body
information_sought	string	What the RTI requested (1 sentence)
exemptions_cited	list[string]	RTI Act sections invoked e.g. ["8(1)(j)"]
outcome	string	See outcome labels below
penalty_inr	float	Penalty amount if imposed
compensation_inr	float	Compensation awarded if any
final_direction	string	Last directive sentence from Commission

Source B — CIC PDF Corpus (cic_annotated)

Extracted from 298 CIC decision PDFs collected from dsscic.nic.in.

Field	Type	Description
filename	string	Original PDF filename
doc_subtype	string	PRIMARY_DECISION / ADJUNCT_COMPLIANCE / FULL_BENCH
doc_format	string	2023a / 2023b / 2026
case_no	string	CIC case number e.g. CIC/CSWRI/A/2021/136051
commissioner	string	Information Commissioner name
public_authority	string	Respondent department
issue	string	Information sought (IRAC Issue component)
application	string	Submissions during hearing (IRAC Application)
rules_cited	list[string]	RTI Act sections referenced
conclusion	string	Commission's decision text (IRAC Conclusion)
outcome	string	See outcome labels below
exemptions_cited	list[string]	Section 8(1)(x) exemptions invoked
rti_filed_on	string	Date RTI application filed
cpio_replied_on	string	Date CPIO replied
first_appeal_on	string	Date first appeal filed
date_of_hearing	string	Date of CIC hearing
adjunct_outcome	string	For ADJUNCT_COMPLIANCE: SCN_DROPPED / PENALTY_IMPOSED etc.

---

Outcome Labels

Label	Description	Count (A)	Count (B)
INFORMATION_DIRECTED	Commission directed disclosure	524	15
APPEAL_DISMISSED	Appeal dismissed, PIO upheld	380	69
UNKNOWN	Requires human review	134	117
PENALTY_IMPOSED	Penalty under Section 20	92	10
PARTIAL_RELIEF	Partial information directed	76	0
COMPLAINT_S18	Section 18 complaint disposed	0	18
REMANDED	Referred back to PIO/FAA	11	5
WITHDRAWN	Appellant withdrew	0	5
ADJOURNED	Adjourned sine die	1	0

---

Exemption Distribution

467 total exemption citations across both sources:

Section	Description	Count
8(1)(j)	Personal information	158
8(1)(d)	Commercial confidence	77
8(1)(e)	Fiduciary relationship	76
8(1)(h)	Impeding investigation	71
8(1)(g)	Life/safety of person	31
8(1)(a)	Sovereignty/security	25
8(1)(i)	Cabinet papers	16
8(1)(c)	Parliament privilege	6
8(1)(b)	Contempt of court	5
8(1)(f)	Fiduciary (foreign govt)	2

---

Benchmark Tasks

RTI-Bench supports four benchmark tasks:

Task 1 — Outcome Prediction Given the background narrative, predict the Commission outcome (multi-class classification). Evaluate with macro-F1. Majority baseline: 44.7% accuracy, 14.3% macro-F1.

Task 2 — Exemption Classification Given narrative and decision text, identify which RTI Act exemptions were invoked (multi-label). Evaluate with micro-F1 and per-section F1.

Task 3 — Compliance Outcome Prediction Given the original CIC directive and respondent's compliance submission, predict compliance outcome (SCN_DROPPED / SCN_CONTINUED / PENALTY_IMPOSED). Uses adjunct decision subset (n=17 in v1.0).

Task 4 — Plain-Language Summarisation Given the full decision text, generate a citizen-accessible summary. Reference summaries available in Source A response field. Evaluate with ROUGE-L, BERTScore, and human faithfulness rubrics.

---

Document Format Generations (Source B)

A notable finding is the evolution of CIC document templates:

• Format 2023a (n=111): O R D E R / Facts / Decision: headings; bilingual Hindi-English headers; party blocks on separate lines.
• Format 2023b (n=21): Observations: + Decision: sections; Date of Decision in header.
• Format 2026 (n=166): DECISION all-caps block; explicit INFORMATION COMMISSIONER: Name label; inline party names; slash-separated dates.

---

Data Collection and Extraction

Source A: Rule-based extraction from jatinmehra/RTI-CASE-DATASET using regex patterns for public authority, exemption sections, outcome language, penalty amounts, and final direction. Runs in ~30 seconds for 1,218 rows.

Source B: CIC decision PDFs collected from dsscic.nic.in (manual download with CAPTCHA). Text extracted using PyMuPDF. Format-aware rule-based extractors applied per document generation. Full pipeline runs in ~90 seconds for 298 PDFs. No LLM annotation used anywhere.

All scripts available at: [GitHub link — add after upload]

---

Commissioners Represented (Source B)

Commissioner	Count
Amita Pandove	75
Jaya Varma Sinha	40
Sudha Rani Relangi	35
Ashutosh Chaturvedi	30
Vanaja N. Sarna	14

---

Limitations

• 17.2% of primary cases carry UNKNOWN outcome labels (49% in CIC PDF subset), reflecting indirect Commission language not captured by current patterns; these require human review before use in supervised classification.
• CIC PDF corpus concentrated among 5 commissioners; may not represent the full diversity of Information Commission adjudication across India.
• Appellant name field has low coverage (<25%) due to format variation; not required for the four benchmark tasks.
• Does not include State Information Commission decisions.
• Exemption extraction relies on explicit section citations; implicit reasoning about exemptions is not captured.

---

Ethical Considerations

All data is drawn from publicly available sources. CIC decisions are public records under Indian law. Personal names of appellants appear as part of the public record; researchers should apply appropriate de-identification for downstream applications involving sensitive matters.

Intended use: improving AI-assisted access to justice for Indian citizens. Not intended for use in discouraging legitimate RTI appeals.

---

Citation

@dataset{bose2025rtibench,
  title     = {RTI-Bench: A Structured Dataset for Indian Right-to-Information Decision Analysis},
  author    = {Bose, Joy},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/joybose/rti-bench}
}

---

License

CC BY 4.0. Source data from CIC portal is public domain under Indian government open data policy. Source A derived from jatinmehra/RTI-CASE-DATASET (original license applies to that subset).