biplaw/rbi-amendment-extraction

RBI amendment-graph extraction: benchmark and training set

Three splits over 506 Reserve Bank of India circulars (February 2024 – August 2026), for extracting what an instrument changes, how, under what legal power, and from when.

Code and scoring: https://github.com/BiplawSingh/beat-the-giant Model trained on this data: https://huggingface.co/biplaw/Llama-3.1-8B-RBI-Amendments-LoRA

split	documents	labels	use
test	101	hand-written, frozen	the benchmark
train	354	model-generated, corrected	fine-tuning
val	51	model-generated, corrected	choosing a checkpoint

Each row is a document and five fields:

{"instrument_type": "Amendment Directions",
 "applies_to": ["Rural Co-operative Banks"],
 "amends": [{"instrument_title": "...", "year": 2025, "relation": "amends"}],
 "commencement": {"mode": "dated", "date": "2027-01-01"},
 "legal_authority": [{"act": "Banking Regulation Act, 1949", "sections": ["21", "35A"]}]}

The split is the point

74% of this corpus sits in families of near-identical documents. RBI issues the same rule once per class of institution, so one amendment exists as up to ten documents differing by a noun phrase. A random split puts near-copies on both sides, and a model trained that way scores brilliantly for the wrong reason.

Whole families are assigned to one side. The invariant is asserted by a test on every push. If you re-split this data, split by family or the numbers mean nothing.

How the labels were made

Test (101). Regex drafts filled the mechanical fields; every document was then read and corrected by hand, and amends — the hardest field, and the one that decides the headline — was written from scratch. The key was frozen before any model was run against it. Four labels changed afterwards; each is recorded with the quoted text, the before/after scores, and, where model disagreement prompted the re-read, that disclosure.

Train (354) and val (51). Labelled by openai/gpt-oss-120b, then corrected to a written guide: 35 by mechanical rule, 9 by hand. Applying the same rules to the teacher's own test answers raises its score from 0.837 to 0.926, which is the measurable estimate of what the correction is worth.

Raw teacher output is published alongside the corrected labels, and every changed label records which rule changed it.

Benchmark results

model	HARD F1	exact
gpt-5.4-mini + guide in prompt	0.941	86/101
Llama-3.1-8B QLoRA fine-tune	0.935	67/101
gpt-4o-mini	0.843	58/101
openai/gpt-oss-120b	0.837	60/101
gpt-5.4-mini	0.833	69/101
untuned Llama-3.1-8B-Instruct	0.000	0/101

Scoring: exact match plus field-level F1, with the headline restricted to the hard fields. A parse failure scores zero and stays in the denominator. Scoring code, the frozen key and every run's raw answers are in the repository.

Known characteristics

• One test document carries ~30% of the amends evidence (a 38-instrument supersession list). Report scores with and without it.
• Rare relations are rare. Across the training labels: amends 270, supersedes 35, consolidates 14, repeals 8. Two wrong supersedes labels caused 37 relation errors in a fine-tune until they were fixed.
• Every model sees a window, not the whole document: first 4,800 and last 1,600 characters. Within it the test split retains 76/76 commencement clauses, 70/70 authority clauses and 47/48 amendment clauses.
• Training and validation labels are model-generated. They are good enough to train on and not a ground truth to measure against. Only the test split is hand-written.

Source

Reserve Bank of India circulars, retrieved 2026. RBI publishes these documents publicly; the corpus is included because it cannot be re-harvested (the host is WAF-blocked) and every number was measured on this exact snapshot. Labels and code are released CC BY 4.0; the underlying documents remain RBI's.