sinhal/indian-sc-outcome-bench

Indian Supreme Court Judgment Outcome Bench

380 Indian Supreme Court judgments. The task: predict whether the appeal or petition was allowed or dismissed, given only the opening of the judgment (facts and arguments, with the decision removed).

The set is built to separate reasoning from recall:

• Two halves with the same task and format: 190 recent judgments (2023–24) and 190 old ones (1995–2005), both with the same 135 allowed / 55 dismissed split.

• A second, name-only condition: each case is also asked with only its title and date. A model that does better on old cases, or that can answer from the name alone, is remembering, not reasoning.

• Code and raw predictions: github.com/Zephyr4772/indian-sc-outcome-bench

• Source judgments: sinhal/Indian_Supreme_Court_Judgments

Results: five open 7–8B models (zero-shot)

Balanced accuracy is the headline metric: chance and always-the-same-answer both score 50%. The raw majority baseline (always "allowed") is 71.1%.

With judgment text:

model	bal. acc	old	recent	old − recent (95% CI)	answers "allowed"	AUROC
Mistral-7B-Instruct-v0.3	55.4	54.2	56.5	−2.3 [−12.0, +7.7]	75.5%	n/a*
Qwen3-8B (no thinking)	52.1	51.5	52.7	−1.1 [−12.0, +9.7]	42.1%	0.58
Qwen2.5-7B-Instruct	51.6	49.7	53.6	−4.0 [−15.3, +6.8]	40.5%	0.58
Phi-4-mini-instruct	51.1	51.4	50.8	+0.7 [−6.3, +7.5]	11.6%	0.55
Granite-3.3-8B-instruct	49.4	49.4	49.3	+0.1 [−6.6, +7.0]	9.2%	0.59

Name and date only (recall probe):

model	bal. acc	old	recent	answers "allowed"	unparsed
Qwen3-8B	50.4	50.3	50.5	84.2%	0%
Qwen2.5-7B-Instruct	50.0	50.0	50.0	100%	0%
Phi-4-mini-instruct	49.3	50.0	48.6	0.8%	0%
Granite-3.3-8B-instruct	47.1	42.5	51.8	26.3%	1.6%
Mistral-7B-Instruct-v0.3	28.7	50.0	7.4	57.9%	42.1% (recent cases only, see below)

*Mistral's first generated token is a bare space, so its option probabilities were not captured. Its answers are valid.

Findings:

1. All five models are at chance. Balanced accuracy is 49–55%, and the probability signal is faint (AUROC ≤ 0.59).
2. Each model mostly applies a fixed bias instead of reading the case. Granite and Phi answer "dismissed" ~90% of the time, while Mistral answers "allowed" 76% of the time. The true split is 71% / 29%.
3. No memorisation is visible at this size. Old and recent cases score the same for every model (every gap's confidence interval includes 0), and no model recalls outcomes from the case name.
4. Mistral recognises dates beyond its training data. Given only the name and date, it declined to answer for 160 of 190 recent cases ("Pending…", "I'm an AI…") and for 0 of 190 old ones. On old cases it is still at chance, so this is date-awareness, not recall.

Setup: zero-shot, greedy decoding, the same prompt for every model, vLLM 0.31 on Kaggle 2× T4 (fp16), max context 10,240 tokens. The longest prompt is 8,415 tokens with the Granite tokenizer, so nothing is truncated.

How items were built (build.py)

• Source: English, full-length judgments (not short orders) from 1995–2005 and 2023–24. Duplicate PDFs listed under several years are removed.
• Label: kept only when the dataset's outcome field and the judgment's own closing paragraphs agree ("…appeal is allowed" / "…dismissed"). Partly-allowed, disposed-of and ambiguous cases are excluded. The metadata outcome contradicted the judgment text in 8% of old and 3% of recent candidates, which is why the cross-check matters.
• Input: page headers and signatures are stripped, then the text is cut at a paragraph boundary before the final 35%, capped at 14,000 characters. Any item whose kept text still states the outcome is dropped.
• Pairing: all 190 recent items that pass, plus a seeded random sample of old items with the same label counts.

Fields

field	description
id	ISC-0000 …
era	recent (2023–24) or old (1995–2005)
year, judgment_date	decision date
petitioner, respondent	party names (used by the name-only probe)
text	opening of the judgment, decision removed (median ~10.5k characters)
label	allowed / dismissed
source_id, pdf, full_text_chars	link back to the source dataset

Limitations

• Small: 380 items, and the dismissed class has 110. Old − recent gaps under about ±10 points are within noise; the confidence intervals are reported for that reason.
• Binary labels only. Partly-allowed and disposed-of outcomes are excluded.
• The input can still lean toward the outcome. The opening may summarise the court's reasoning. Explicit outcome statements are filtered, but the input isn't a sanitised fact pattern.
• "Recent" is relative. 2023 cases may be inside newer models' training data. The name-only probe is the direct check.
• English only, because the source's vernacular texts have broken font encoding.

Load

from datasets import load_dataset

ds = load_dataset("sinhal/indian-sc-outcome-bench", split="test")

Submit your model's results in a discussion and I'll add it to the table.