Requirements
- Target platform
- OpenClaw
- Install method
- Manual import
- Extraction
- Extract archive
- Prerequisites
- OpenClaw
- Primary doc
- SKILL.md
Tencent SkillHub Β· AI
Llm As Judge
Build a cost-efficient LLM evaluation ensemble with sampling, tiebreakers, and deterministic validators. Learned from 600+ production runs judging local Olla...
skill openclawclawhub Free
Build a cost-efficient LLM evaluation ensemble with sampling, tiebreakers, and deterministic validators. Learned from 600+ production runs judging local Olla...
β¬ 0 downloads β
0 stars Unverified but indexed
Install for OpenClaw
Quick setup
- Download the package from Yavira.
- Extract the archive and review SKILL.md first.
- Import or place the package into your OpenClaw setup.
Package facts
- Download mode
- Yavira redirect
- Package format
- ZIP package
- Source platform
- Tencent SkillHub
- What's included
- SKILL.md
Validation
- Use the Yavira download entry.
- Review SKILL.md after the package is downloaded.
- Confirm the extracted package contains the expected setup assets.
Install with your agent
Agent handoff
Hand the extracted package to your coding agent with a concrete install brief instead of figuring it out manually.
- Download the package from Yavira.
- Extract it into a folder your agent can access.
- Paste one of the prompts below and point your agent at the extracted folder.
New install
I downloaded a skill package from Yavira. Read SKILL.md from the extracted folder and install it by following the included instructions. Tell me what you changed and call out any manual steps you could not complete.
Upgrade existing
I downloaded an updated skill package from Yavira. Read SKILL.md from the extracted folder, compare it with my current installation, and upgrade it while preserving any custom configuration unless the package docs explicitly say otherwise. Summarize what changed and any follow-up checks I should run.
Trust & source
Release facts
- Source
- Tencent SkillHub
- Verification
- Indexed source record
- Version
- 1.0.1
Provenance
- Publisher
- nissan
- Source page
- View original listing
- Canonical URL
- Open canonical page
Documentation
LLM-as-Judge
Build a cost-efficient LLM evaluation ensemble for comparing and scoring generative AI outputs at scale.
When to Use
Evaluating generative AI outputs across multiple models at scale (100+ runs) Comparing local/OSS models against cloud baselines in shadow-testing pipelines Building promotion gates where models must prove quality before serving production traffic Any scenario where deterministic tests alone can't capture output quality
When NOT to Use
One-off evaluations (just read the output yourself) Tasks with deterministic correct answers (use exact-match or unit tests) When you can't afford any external API calls (this pattern uses Claude/GPT as judges)
Layer 1: Deterministic Validators (Free, Instant)
Run on 100% of outputs. Zero cost. Catches obvious failures before burning judge tokens. JSON schema validation β does the output parse? Does it match the expected schema? Regex checks β required fields present, format constraints met Length bounds β output within acceptable min/max character count Entity presence β do required entities from the input appear in the output? If Layer 1 fails, score is 0.0 β no need to invoke expensive judges.
Layer 2: Heuristic Drift Detection (Cheap, Fast)
Run on 100% of outputs that pass Layer 1. Minimal cost (local computation only). Entity overlap β what fraction of entities in the ground truth appear in the candidate? Numerical consistency β do numbers in the output match source data? Novel fact detection β does the output introduce facts not present in the input/context? Novel facts suggest hallucination. Structural similarity β does the output follow the same structural pattern as ground truth? Layer 2 produces heuristic scores (0.0β1.0) that contribute to the final weighted score.
Layer 3: LLM Judges (Expensive, High Quality)
Sampled at 15% of runs to control cost. Forced to 100% during promotion gates. Two independent judges (e.g., Claude + GPT-4o) score the output. Each judge evaluates all 6 dimensions independently. Tiebreaker pattern: When primary judges disagree by Ξ β₯ 0.20 on any dimension, a third judge is invoked. The tiebreaker score replaces the outlier. This reduced score variance by 34% at only 8% additional cost.
The 6 Scoring Dimensions
DimensionWeightWhat It MeasuresStructural accuracy0.20Format compliance, schema adherenceSemantic similarity0.25Meaning preservation vs ground truthFactual accuracy0.25Correctness of facts, numbers, entitiesTask completion0.15Does it actually answer the question?Tool use correctness0.05Valid tool calls (when applicable)Latency0.10Response time within acceptable bounds Weights are configurable per task type. Tool use weight is redistributed when not applicable.
Critical Lesson: None β 0.0
When a dimension is not sampled (LLM judge not invoked on this run), record the score as null, not 0.0. Unsampled dimensions must be excluded from the weighted average, not treated as failures. Early bug: recording unsampled dimensions as 0.0 created a systematic 0.03β0.08 downward bias across all models. The fix: null means "not measured", which is fundamentally different from "scored zero". # WRONG β penalises unsampled dimensions weighted = sum(s * w for s, w in zip(scores, weights)) / sum(weights) # RIGHT β exclude null dimensions pairs = [(s, w) for s, w in zip(scores, weights) if s is not None] weighted = sum(s * w for s, w in pairs) / sum(w for _, w in pairs)
Cost Estimate
With 15% LLM sampling, average cost per evaluated run: ~$0.003 Layer 1 + Layer 2: $0.00 (local computation) Layer 3 (15% of runs): ~$0.02 per judged run Γ 0.15 = ~$0.003 Tiebreaker (fires ~12% of judged runs): adds ~$0.0003 At 200 runs for promotion: total judge cost β $0.60 per model per task type.
Worked Example: Summarisation Evaluation
from evaluation import JudgeEnsemble, DeterministicValidator, HeuristicScorer # Layer 1: must be valid text, 50-500 chars validator = DeterministicValidator( min_length=50, max_length=500, required_format="text", ) # Layer 2: check entity overlap with source heuristic = HeuristicScorer( check_entity_overlap=True, check_novel_facts=True, check_numerical_consistency=True, ) # Layer 3: LLM judges (sampled) ensemble = JudgeEnsemble( judges=["claude-sonnet-4-20250514", "gpt-4o"], tiebreaker="claude-sonnet-4-20250514", sample_rate=0.15, tiebreaker_threshold=0.20, dimensions=["structural", "semantic", "factual", "completion", "latency"], ) # Evaluate result = ensemble.evaluate( task_type="summarize", ground_truth=gt_response, candidate=candidate_response, source_text=original_text, validator=validator, heuristic=heuristic, ) print(f"Weighted score: {result.weighted_score:.3f}") print(f"Dimensions: {result.scores}") # {semantic: 0.95, factual: 0.88, ...} # None values for unsampled dimensions
Tips
Start with Layer 1 β you'd be surprised how many outputs fail basic validation Log everything β store raw judge responses for debugging score disputes Calibrate on 50 runs β before trusting the ensemble, manually review 50 outputs against judge scores Watch for judge drift β LLM judges can be inconsistent across API versions; pin model versions Force judges at gates β 15% sampling is fine for monitoring, but promotion decisions need 100% coverage on the final batch
Category context
Agent frameworks, memory systems, reasoning layers, and model-native orchestration.
Source: Tencent SkillHub
Largest current source with strong distribution and engagement signals.
Package contents
Included in package
1 Docs
- SKILL.md