README.md

metadata

license: apache-2.0
task_categories:
  - text-retrieval
language:
  - en
tags:
  - smart-contracts
  - solidity
  - security
  - evaluation
  - benchmark
size_categories:
  - n<1K

SCAR Evaluation Set

838 held-out contrastive pairs for evaluating smart contract vulnerability retrieval models. Constructed as a 10% stratified holdout from training sources plus the 91 original FORGE-Curated pairs (temporal holdout).

The split is performed at the report level — not the finding level — to prevent information leakage. Hash-based verification (SHA-256 on normalized source) confirms zero code overlap with scar-pairs.

Source Breakdown

Source	Pairs
Solodit	228
msc-audits-with-reasons	178
msc-smart-contract-auditing	131
FORGE-Curated (original)	91
DeFiHackLabs	67
FORGE-Artifacts	59
FORGE-Curated-v2	36
EVuLLM	32
SmartBugs-Curated	14
GitmateAI	2
Total	838

Headline Results on this Eval Set

Method	R@1	R@10	nDCG@10	MRR
BM25	0.504	0.689	0.591	0.566
E5-base-v2	0.456	0.656	0.554	0.530
SPLADE (Qwen)	0.809	0.963	0.892	0.869
SCAR (25 epochs)	0.894	0.977	0.939	0.927
SCAR + BM25 hybrid	0.825	0.983	0.908	0.884

Improvement over BM25 statistically significant at p < 0.0001 (paired bootstrap, n = 10,000).

Usage

from datasets import load_dataset
ds = load_dataset("Farseen0/scar-eval", split="train")

Related

• 🗂️ All SCAR artifacts (Collection)
• scar-pairs — 7,552 training pairs
• scar-corpus — 231k contract corpus
• scar-weights — trained model weights

Paper

This dataset accompanies SCAR: Sparse Code Audit Retriever via SAE-LoRA Adaptation (Farseen Shaikh, 2026).

• Paper: OpenReview submission (under review at EMNLP 2026, ACL ARR March cycle)
• Code: github.com/FarseenSh/scar-retrieval
• Model: Farseen0/scar-weights

Citation

@inproceedings{shaikh2026scar,
  title  = {SCAR: Sparse Code Audit Retriever via SAE-LoRA Adaptation},
  author = {Shaikh, Farseen},
  year   = {2026},
  note   = {Under review at EMNLP 2026 (ACL ARR March cycle)},
  url    = {https://openreview.net/forum?id=moD8Hxq9hN}
}

License

Apache 2.0 — free for research and commercial use with attribution.