llm-degen-guard

한국어 문서

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):

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

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...?

License

MIT