GitBench (v2): Real-World GitHub Pull Request State & Security Benchmark for AI Coding Agents

GitBench (originally GitState-Bench) is an open evaluation benchmark developed by the Knot AI team (@Knot-Repo) to rigorously evaluate how AI coding assistants, code search systems, and Retrieval-Augmented Generation (RAG) engines handle temporal code evolution, obsolete baseline specifications, unmerged RFC proposals, and 12 families of production credentials.

---

Benchmark Motivation

Most coding benchmarks (e.g., HumanEval, MBPP, SWE-bench) evaluate whether a model can synthesize code from a static snapshot. However, in production engineering environments, developer context evolves continuously through Git commits and Pull Requests.

AI agents and vector retrieval layers regularly fail in three high-severity ways:

1. Temporal Obsolescence (The "Verbose Legacy vs. Concise Update" Trap): Retrieving a verbose legacy architecture specification (superseded_v1) instead of a concise newly merged PR (active_v2_gold) because both share $>85%$ of their domain vocabulary.
2. Brainstorm / Draft RFC Contamination: Treating unmerged draft RFCs or exploratory proposals (brainstorm_unconfirmed) as authoritative production behavior.
3. High-Entropy Secret Leakage vs. False-Positive Over-Quarantine: Leaking unprefixed HMAC keys or authenticated database/Redis DSN URIs into prompts (admin_secret_trap)—or conversely, falsely quarantining benign engineering notes that contain high-entropy Git commit SHAs and UUIDv4 trace IDs.

---

Dataset Scale & Repository Coverage

• Total Corpus Documents: 1,788 (447 merged active specifications, 447 superseded baseline specifications, 447 unmerged RFC/draft proposals, 447 multi-family runtime secret configurations)
• Developer Queries: 447 natural active implementation queries + 447 paired historical audit queries (historical_text)
• Relevance Judgments (qrels): 1,788 graded relevance pairs (4,023 total records across splits)
• 25 Sampled Repositories: fastapi/fastapi, pydantic/pydantic, anthropics/anthropic-sdk-python, vercel/ai, encode/httpx, psf/requests, pallets/flask, tiangolo/sqlmodel, openai/openai-python, langchain-ai/langchain, run-llama/llama_index, astral-sh/uv, huggingface/transformers, vllm-project/vllm, pandas-dev/pandas, numpy/numpy, scikit-learn/scikit-learn, pytorch/pytorch, sqlalchemy/sqlalchemy, celery/celery, pytest-dev/pytest, Textualize/rich, streamlit/streamlit, gradio-app/gradio, BerriAI/litellm.

---

12-Family Credential & Benign High-Entropy Design

Unlike synthetic benchmarks that test only a single hardcoded token prefix (such as ghp_), GitBench v2 distributes 447 secret traps across 12 realistic credential families while embedding benign high-entropy Git commit SHAs and UUIDv4 trace identifiers into normal engineering documents:

• 50% Unprefixed High-Entropy Secrets (6 Families): WEBHOOK_SIGNING_SECRET, JWT_HS256_PRIVATE_KEY, MASTER_ENCRYPTION_KEY, OAUTH_CLIENT_SECRET, INTERNAL_AUTH_TOKEN, SESSION_HMAC_SECRET ($H \ge 3.85\text{ bits/char}$, no vendor prefix—missed completely by prefix regexes).
• 25% Authenticated Service & Database DSN URIs (3 Families): PostgreSQL (postgresql://...), Redis Cluster (redis://:...), and AMQP/Celery (amqp://...) connection strings with high-entropy embedded passwords ($H \ge 3.10\text{ bits/char}$).
• 25% Vendor-Prefixed Tokens & Private Key Headers (3 Families): AWS (AKIA...), Anthropic (sk-ant-...), and OPENSSH PRIVATE KEY blocks.

---

Empirical Baseline Results (all-MiniLM-L6-v2, 384D)

Architecture	Scoped Active v2 (Invariant)	Global Multi-PR Top-1	Historical Audit Recall@1	Credential Leak @ 5	Benign Code False Quarantines	In-Memory Latency
Standard Vector RAG (HNSW Cosine)	60.2%	56.8%	43.0%	65.1%	0 / 1,341	0.013 ms
Hybrid Lexical + Vendor-Prefix Regex Only	63.5%	61.5%	43.0%	49.7% (Misses unprefixed)	0 / 1,341	0.070 ms
Hybrid Lexical + Unconditioned Shannon Entropy	7.6%	8.5%	43.0%	1.3%	1,020 / 1,341 (SHAs/UUIDs)	0.068 ms
Knot CARMA-16 Engine + UCS Protocol	100.0% (0 Stale/Draft)	89.5%	86.4%	0.0% (12/12 Families)	0 / 1,341 (0.0% FPR)	0.079 ms

---

Quickstart

from datasets import load_dataset

# Load GitBench v2 directly from Hugging Face (4,023 records)
corpus  = load_dataset("knot-ai/gitbench", split="corpus")   # 1,788 PR state documents
queries = load_dataset("knot-ai/gitbench", split="queries")  # 447 active + historical queries
qrels   = load_dataset("knot-ai/gitbench", split="qrels")    # 1,788 TREC/BEIR relevance judgments

---

Citation & Attribution

@misc{knot_ai_gitbench_2026,
  author = {Knot AI Team},
  title = {GitBench: Evaluating Temporal Causal Freshness and Secret Containment in AI Coding Agents},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/knot-ai/gitbench}}
}