- name
SKILL.md
Le fichier du skill reste dans le dépôt de l'éditeur.
Collez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add agentscope-ai/OpenJudge --skill 08-bootstrap -a cursor -yCollez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add agentscope-ai/OpenJudge --skill 08-bootstrap -a claude-code -yCollez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add agentscope-ai/OpenJudge --skill 08-bootstrap -a codex -yCollez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add agentscope-ai/OpenJudge --skill 08-bootstrap -a grok -yEnvoyez le message copié dans un chat Grok Bot. Le Bot récupère ce SKILL.md sur GitHub et l'ajoute aux skills que partagent tous vos Bots.
Install this skill: https://github.com/agentscope-ai/OpenJudge/tree/HEAD/skills/eval_pipeline/08-bootstrapDe l'éditeur
Cold-start an evaluation system when you have nothing. In 30 minutes you get a working v0 grader and a clear path to a calibrated, trustworthy evaluation.
Requires OpenJudge (
pip install py-openjudge) for the grader generators (SimpleRubricsGenerator/IterativeRubricsGenerator). The interview, stratification, and calibration-roadmap methodology is SDK-independent.
You MUST create a task for each item and complete them in order:
Ask the user to describe their system in one go:
To bootstrap your evaluation, I need to understand what you're building.
Please describe (all at once):
- What does your system do? Who uses it?
- What are 3 examples of perfect outputs?
- What are 3 things the system must never do?
- What failures worry you most?
Don't drip-feed these questions. One prompt, one answer. If the user provides a spec doc or design document instead, read that directly.
Use OpenJudge's SimpleRubricsGenerator to create a zero-shot grader from the
product description:
import asyncio
from openjudge.models.openai_chat_model import OpenAIChatModel
from openjudge.generator.simple_rubric.generator import (
SimpleRubricsGenerator,
SimpleRubricsGeneratorConfig,
)
from openjudge.runner.grading_runner import GradingRunner
# OpenAIChatModel reads OPENAI_API_KEY / OPENAI_BASE_URL from the environment.
# For Aliyun DashScope (Bailian): set OPENAI_BASE_URL to
# https://dashscope.aliyuncs.com/compatible-mode/v1 and OPENAI_API_KEY to your key.
model = OpenAIChatModel(model="qwen-plus") # or "gpt-4o", etc.
config = SimpleRubricsGeneratorConfig(
grader_name="Initial Quality Grader",
model=model,
task_description="<summarize from the interview>",
scenario="<usage context from interview>",
min_score=0,
max_score=1,
)
generator = SimpleRubricsGenerator(config)
grader = await generator.generate(
dataset=[],
sample_queries=[
"<example query 1 from interview>",
"<example query 2 from interview>",
"<example query 3 from interview>",
],
)
Why zero-shot instead of asking the user to write criteria? At this stage, the user doesn't know what "good" means operationally. The generator produces a reasonable starting point. The user refines it after seeing v0 results.
Generate 30 test inputs with stratification. Use 3 different prompt templates for diversity:
Template 1: "Generate a typical {domain} query for a {user_type}"
Template 2: "Create an ambiguous {domain} query where intent is unclear"
Template 3: "Generate an edge-case {domain} query that's unusual but realistic"
Target distribution:
Critical: Generate inputs ONLY. Never generate labels. The labels come from running the actual system and getting human judgments.
# The dataset format for GradingRunner
dataset = [
{
"query": "What's the status of my order #12345?",
"response": "<will be filled by running the system>",
},
# ... 30 inputs
]
Plug the generated grader into GradingRunner:
from openjudge.runner.grading_runner import GradingRunner
from openjudge.graders.schema import GraderScore, GraderError
runner = GradingRunner(
grader_configs={"v0_quality": grader},
max_concurrency=8,
)
results = await runner.arun(dataset)
scores = [r.score for r in results["v0_quality"] if isinstance(r, GraderScore)]
errors = [r for r in results["v0_quality"] if isinstance(r, GraderError)]
print(f"V0 Results: avg={sum(scores)/len(scores):.2f}, errors={len(errors)}")
The v0 grader is uncalibrated — you don't know its TPR/TNR yet. Give the user an exact path to trustworthiness:
Your v0 evaluation is ready. Here's the path to a calibrated system:
Phase 1 (now): Run the v0 grader on 30 inputs to get a baseline.
→ The grader is UNCALIBRATED. Treat scores as directional, not definitive.
Phase 2 (1-2 weeks): Collect 50 human-labeled examples (25 pass + 25 fail).
→ For each system output, have a human mark pass/fail against the criterion.
→ Store labels in labels/<grader_name>.jsonl
Phase 3: When you have 50 labels, run 03-align-human to:
→ Measure TPR/TNR of the v0 grader
→ Detect biases (position, verbosity, self-enhancement)
→ Get a human-reduction roadmap
Phase 4: When TPR >= 0.8 and TNR >= 0.8:
→ The grader is calibrated and can be used as a production gate
IterativeRubricsGenerator
instead of SimpleRubricsGenerator for data-driven grader creation:from openjudge.generator.iterative_rubric.generator import (
IterativeRubricsGenerator,
IterativePointwiseRubricsGeneratorConfig,
)
config = IterativePointwiseRubricsGeneratorConfig(
grader_name="Data-Driven Grader",
model=model,
task_description="<from interview>",
min_score=0, max_score=1,
max_epochs=3,
batch_size=10,
)
generator = IterativeRubricsGenerator(config)
grader = await generator.generate(dataset=labeled_data) # 20+ labeled examples
After 08-bootstrap:
03-align-human: Once 50 human labels are collected, calibrate the grader.01-eval-design: If you want a properly stratified dataset beyond the v0 30 inputs.02-metric-design: If you need multiple graders for different dimensions.>
---name: bootstrapdescription: >---# skills/eval_pipeline/08-bootstrap/SKILL.mdFAQ
Téléchargez skills/eval_pipeline/08-bootstrap/SKILL.md auprès de l'éditeur, ou copiez le chemin et ouvrez le dépôt. PlugMCP répertorie le skill et n'héberge pas le fichier.
L'éditeur le maintient sur https://github.com/agentscope-ai/OpenJudge.
Claude, Cursor, Codex et Grok Build peuvent lire un SKILL.md que vous ajoutez depuis le dépôt de l'éditeur. Grok Bot l'installe lorsque vous envoyez dans un chat le lien du dossier GitHub du skill.