dsh-llm-verifier
Best-of-3/5 orchestration and LLM-as-a-Verifier selection for DeepSeek Harness
AI Analysis
核心用途是通过多候选响应与大模型验证器机制提升输出质量。适合对生成结果准确性要求极高的复杂任务。必要条件包括 DSH 0.1.0-rc.7、Node 24、pnpm 及 uv 工具。
Install
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:Web0926/dsh-llm-verifierREADME
Read the full README ↗Configuration
Default values:
| Setting | Default |
|---|---|
defaultCandidateCount | 3 |
candidateProfile | headless |
credentialRef | DEEPSEEK_API_KEY |
verifierModel | deepseek-v4-flash |
nEvaluations | 2 |
maxVerifierWorkers | 8 |
verifierEffort | high |
verifierMaxTokens | 32768 |
candidateTimeoutMs | 1200000 |
validationTimeoutMs | 600000 |
runTimeoutMs | 2700000 |
maxVerifierTraceBytes | 524288 |
stateDirectory | $DSH_HOME/llm-verifier |
Override the plugin entry in the Web profile's cordis.patch.yml:
- id: llm-verifier
config:
defaultCandidateCount: 3
candidateProfile: headless
credentialRef: DEEPSEEK_API_KEY
verifierModel: deepseek-v4-flash
nEvaluations: 2
maxVerifierWorkers: 8
verifierEffort: high
verifierMaxTokens: 32768
candidateTimeoutMs: 1200000
validationTimeoutMs: 600000
runTimeoutMs: 2700000
maxVerifierTraceBytes: 524288
stateDirectory: $DSH_HOME/llm-verifier
nEvaluations accepts 1–4, maxVerifierWorkers accepts 1–16, and verifierEffort accepts low, high, or max. The verifier model name must begin with deepseek-.