llm-degen-guard

한국어 문서 · llms.txt

Model-agnostic LLM output degeneration detector — 4-signal composite scoring in a single pass.

Install

pip install llm-degen-guard

Quick Start

from degen_guard import DegenGuard

# Batch check
report = DegenGuard.check("Some LLM output text...")
print(report.is_degenerate, report.composite_score)

# Streaming check
guard = DegenGuard()
for chunk in llm_stream:
    is_degen, score = guard.feed(chunk)
    if is_degen:
        print(f"Degeneration detected! score={score:.2f}")
        break

How It Works

Four independent signals are combined into a single composite score (0.0 = normal, 1.0 = degenerate):

Signal	Weight	What it measures
N-gram diversity	0.35	Distinct 3-gram ratio — low diversity = repetitive
Compression ratio	0.30	zlib compression ratio — highly compressible = repetitive
Substring match	0.20	Exact overlap between first/second half of window
Line diversity	0.15	Unique line ratio — duplicate lines = degenerate

Structural content (code, lists, tables) gets a discount factor to reduce false positives.

flowchart TD
    A["📝 LLM Output"] --> B["N-gram Diversity\nweight: 0.35"]
    A --> C["Compression Ratio\nweight: 0.30"]
    A --> D["Substring Match\nweight: 0.20"]
    A --> E["Line Diversity\nweight: 0.15"]
    B --> F["Weighted\nComposite Score"]
    C --> F
    D --> F
    E --> F
    F --> G["Structural\nDiscount"]
    G --> H["Final Score\n0.0 — 1.0"]

Loading

Streaming API

guard = DegenGuard(
    window_size=256,       # characters to analyze
    check_interval=64,     # characters between checks
    alert_threshold=3,     # consecutive alerts before flagging
    score_threshold=0.50,  # minimum score to trigger alert
)

for chunk in llm_stream:
    is_degen, score = guard.feed(chunk)
    if is_degen:
        break

guard.reset()  # reuse for next stream

Batch API

report = DegenGuard.check(full_text)

report.is_degenerate      # bool
report.composite_score    # 0.0 ~ 1.0
report.signals            # tuple of SignalDetail
report.structural_penalty # < 1.0 if code/tables detected
report.text_length        # input length

Why not...?

Alternative	Limitation	llm-degen-guard
Antislop Sampler	Needs model weight access (no API support)	Works on any output text
SpecRA	Paper only, no code available	pip install ready
UQLM	Requires N generations (cost × N)	Single pass, real-time streaming
repetition_penalty	Penalizes prompt tokens, causes incoherence	Post-generation, no side effects

Used in

• watchdeck — Web page monitoring with visual diffs and safety guards

License

MIT

---

_{Part of the QuartzUnit ecosystem — composable Python libraries for data collection, extraction, search, and AI agent safety.}