llm-evaluation-guide
安装方式
先用 Prompt 核对来源,也可切换到安装命令或下载可用的安装包。
用 AI 助手安装
复制提示词,让助手先核对固定版本和依赖。
查看 Prompt收起 Prompt
先检查当前 Agent 是否已安装 SkillPI Skill Finder。如果尚未安装,请按官方仓库的说明安装并启用:https://github.com/Findskills/skillpi。然后使用 Skill Finder 打开此 SkillPI 详情页并定位 "llm-evaluation-guide":https://skillpi.com/zh/skills/deepskill-731c640833c76bf234c6dd00e634919588493143a0c0ca5ce848dcfcfb3cf076。安装前核对上游来源和版本:https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills/blob/bfa0a6e5a843babbf45a1ef1663fd5dc4f07c0ba/skills/43-wentorai-research-plugins/skills/domains/ai-ml/llm-evaluation-guide/SKILL.md。来源不一致时停止并说明差异。核对通过后,用 Skill Finder 安装目标 Skill,并报告安装路径和来源版本。
此命令从 GitHub 来源仓库获取指定技能。运行前请先查看源文件。
npx skills add https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills/tree/bfa0a6e5a843babbf45a1ef1663fd5dc4f07c0ba/skills/43-wentorai-research-plugins/skills/domains/ai-ml/llm-evaluation-guide --skill llm-evaluation-guide来源与配置
查看 GitHub 源码brycewang-stanford/Auto-Empirical-Research-Skills固定提交 bfa0a6e5a843babbf45a1ef1663fd5dc4f07c0ba
配置方式待核实
请查看原始来源中的适用工具、依赖和权限要求。
安装实测待补充
本条目面向文献与学术工具场景,提供“llm-evaluation-guide”技能的目录入口。公开标签:learning、大模型、评估指标、基准测试。请通过固定上游版本核对实际功能、依赖与权限;此页为 Skillπ 独立编写的用途介绍和原站链接,不代表运行安全保证。原文与附件依各自许可开放。
文件浏览
---
name: llm-evaluation-guide
description: "Evaluate and benchmark large language models for research applications"
metadata:
openclaw:
emoji: "🧠"
category: "domains"
subcategory: "ai-ml"
keywords: ["LLM evaluation", "benchmarking", "language models", "model evaluation", "NLP metrics", "BLEU", "perplexity"]
source: "wentor-research-plugins"
---
# LLM Evaluation Guide
A skill for evaluating and benchmarking large language models (LLMs) in research settings. Covers automatic metrics, human evaluation protocols, benchmark suites, evaluation pitfalls, and best practices for reporting LLM performance.
## Evaluation Taxonomy
### Types of Evaluation
```
1. Intrinsic evaluation:
Measures model quality on its own terms
- Perplexity, likelihood, calibration
- Useful for comparing architectures and training procedures
2. Extrinsic evaluation:
Measures model quality on downstream tasks
- Task-specific benchmarks (QA, summarization, classification)
- Closer to real-world usefulness
3. Human evaluation:
Human judges rate model outputs
- Fluency, correctness, helpfulness, safety
- Gold standard but expensive and slow
```
## Automatic Metrics
### Common Metrics by Task
| Task | Metric | Description |
|------|--------|-------------|
| Language modeling | Perplexity | Lower is better; measures prediction quality |
| Machine translation | BLEU, COMET | N-gram overlap; learned quality estimation |
| Summarization | ROUGE-1/2/L | Recall of n-grams against reference |
| Question answering | Exact Match, F1 | Token-level match against reference answer |
| Classification | Accuracy, F1 | Standard classification metrics |
| Generation quality | BERTScore | Semantic similarity via embeddings |
| Factuality | FActScore | Proportion of atomic facts supported by evidence |
### Computing Key Metrics
```python
from collections import Counter
import math
def compute_bleu(reference: list[str], hypothesis: list[str],
max_n: int = 4) -> float:
"""
Compute corpus-level BLEU score (simplified).
Args:
reference: List of reference token sequences
hypothesis: List of hypothesis token sequences
max_n: Maximum n-gram order
"""
precisions = []
for n in range(1, max_n + 1):
num = 0
den = 0
for ref_tokens, hyp_tokens in zip(reference, hypothesis):
ref_ngrams = Counter(
tuple(ref_tokens[i:i+n]) for i in range(len(ref_tokens) - n + 1)
)
hyp_ngrams = Counter(
tuple(hyp_tokens[i:i+n]) for i in range(len(hyp_tokens) - n + 1)
)
clipped = {ng: min(c, ref_ngrams.get(ng, 0))
for ng, c in hyp_ngrams.items()}
num += sum(clipped.values())
den += max(sum(hyp_ngrams.values()), 1)
precisions.append(num / max(den, 1))
# Brevity penalty
ref_len = sum(len(r) for r in reference)
hyp_len = sum(len(h) for h in hypothesis)
bp = math.exp(1 - ref_len / max(hyp_len, 1)) if hyp_len < ref_len else 1.0
# Geometric mean of precisions
log_avg = sum(math.log(max(p, 1e-10)) for p in precisions) / max_n
return bp * math.exp(log_avg)
```
## Benchmark Suites
### Major LLM Benchmarks
```
General knowledge and reasoning:
- MMLU (Massive Multitask Language Understanding): 57 subjects, MCQ
- HellaSwag: Commonsense sentence completion
- ARC (AI2 Reasoning Challenge): Science questions
- WinoGrande: Coreference resolution / commonsense
Coding:
- HumanEval: Python function completion (pass@k)
- MBPP: Mostly basic Python problems
- SWE-bench: Real-world software engineering tasks
Math:
- GSM8K: Grade school math word problems
- MATH: Competition-level mathematics
Safety and alignment:
- TruthfulQA: Resistance to common misconceptions
- BBQ (Bias Benchmark for QA): Social bias in QA
- RealToxicityPrompts: Tendency to generate toxic text
Instruction following:
- MT-Bench: Multi-turn conversation quality (LLM-as-judge)
- AlpacaEval: Instruction-following quality
- Chatbot Arena: ELO-based human preference ranking
```
## Human Evaluation
### Designing a Human Evaluation Protocol
```python
def design_human_eval(task: str, n_annotators: int = 3,
n_examples: int = 200) -> dict:
"""
Design a human evaluation protocol for LLM outputs.
Args:
task: The task being evaluated
n_annotators: Number of independent annotators per example
n_examples: Number of examples to evaluate
"""
return {
"task": task,
"n_annotators": n_annotators,
"n_examples": n_examples,
"criteria": [
{"name": "Fluency", "scale": "1-5",
"description": "Is the text grammatically correct and natural?"},
{"name": "Relevance", "scale": "1-5",
"description": "Does the output address the input/question?"},
{"name": "Correctness", "scale": "1-5",
"description": "Is the factual content accurate?"},
{"name": "Helpfulness", "scale": "1-5",
"description": "Would a user find this response useful?"}
],
"agreement_metric": "Krippendorff's alpha (ordinal)",
"presentation": "Randomize model order; blind annotators to model identity",
"calibration": "Have all annotators rate 20 shared examples first",
"cost_estimate": f"~{n_examples * n_annotators * 0.50:.0f} USD at typical rates"
}
```
## Evaluation Pitfalls
### Common Mistakes
```
1. Data contamination:
Test data may appear in the LLM's training set.
Mitigation: Use held-out datasets, check for contamination,
create new test sets.
2. Metric gaming:
High BLEU does not mean high quality; ROUGE rewards verbosity.
Mitigation: Use multiple metrics and human evaluation.
3. Cherry-picking examples:
Showing only best-case outputs misrepresents model capabilities.
Mitigation: Report aggregate metrics over full test sets.
4. Ignoring variance:
LLM outputs vary with temperature and random seeds.
Mitigation: Report mean and standard deviation over multiple runs.
5. Unfair comparisons:
Comparing models with different prompt formats or few-shot counts.
Mitigation: Standardize prompts and report all hyperparameters.
```
## Reporting Standards
When publishing LLM evaluation results, report: model name and version, parameter count and architecture, evaluation dataset with version number, exact prompts used (include in appendix), number of few-shot examples, decoding parameters (temperature, top-p, max tokens), multiple metrics (not just one), confidence intervals or significance tests, and hardware and inference cost where relevant.
版本历史
-
v2026.07
GitHub ↗
原版更新说明
## Highlights - **First tagged release.** Every gate is green: `make check` (validate + 183 unit tests + eval-harness + numeric benchmarks), validate-catalog, quality-evals, and OpenSSF Scorecard on `main`. - **Rigor double loop at 13 method families.** Each family closes taxonomy tag → eval scenario (does the agent avoid the method's known traps?) → deterministic numeric benchmark (does it recover the true effect?). New this cycle: heterogeneous treatment effects (CATE) and quantile treatment effects (QTE). - **Paper-WorkFlow flagship at 29/29 executable rigor gates**, now merged to its `main` with a drift-gated README badge — the credibility artifact is checked by CI, so it can never silently go stale ([RIGOR.md](https://github.com/brycewang-stanford/Paper-WorkFlow/blob/main/RIGOR.md)). - **1,150 vendored skills / 69 collections**, all license-audited, security-scanned, tracked in a machine-readable catalog with CI-enforced freshness, and importable as a single whole-repo router skill. - **Hardened weekly upstream syncs**: synced trees must pass the full gate before merge; this cycle both sync PRs were verified, patched (StatsPAI's SkillOpt execution-gate card restored), and merged. - **First-party plugin marketplace** (Python / R / Stata empirical-analysis plugins + guide) shipped in the previous cycle, now under a tagged version. ## Known Follow-Ups - Ecosystem benchmark comparison (AERS vs econometrics agents on the shared benchmark suite) — [PLAN-2026-07 week 3](https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills/blob/main/docs/PLAN-2026-07.md). - Decomposition (Oaxaca–Blinder) and shift-share (Bartik) method families (weeks 3–4). - GitHub Pages catalog site, README demo screenshots, five-language README consistency script (week 4). ## Catalog - Top-level collections: **69** - `SKILL.md` files: **1150** - Cataloged agent tools: **335** ([`tools/CATALOG.md`](../tools/CATALOG.md)) ## Methodological rigor - Method families in the coverage map: **13** ([`RIGOR_COVERAGE.md`](RIGOR_COVERAGE.md)) - Eval scenarios: **26** (138 rubric items, 122 auto-checkable) ([`../eval-harness/`](../eval-harness/README.md)) - Benchmark tasks: **13** (70 gold checks, 58 required) ([`../benchmark/`](../benchmark/README.md)) ### Method families (skills tagged) - `iv`: 27 - `rdd`: 25 - `panel-fe`: 19 - `staggered-did`: 17 - `bayesian`: 13 - `did`: 10 - `event-study`: 10 - `matching`: 10 - `synthetic-control`: 10 - `cate`: 8 - `dml`: 8 - `survival`: 3 - `quantile`: 2 ### Eval scenarios by category - `causal-identification`: 14 - `citation-hygiene`: 1 - `reproducibility`: 2 - `research-integrity`: 1 - `runtime-safety`: 1 - `statistical-validity`: 2 - `writing-compliance`: 3 - `writing-style`: 2 ## License buckets - AGPL-3.0: 1 - Apache-2.0: 1 - CC-BY-NC-4.0 (non-commercial): 1 - CC-BY-SA-4.0: 1 - CC-BY-SA-4.0 (repository default): 4 - GPL-3.0: 1 - MIT: 35 - MIT Non-Commercial: 1 - Mixed (per-skill; see folder): 2 - UNKNOWN - check upstream: 22 ## Quality gates - `make check` — catalog/provenance/audit/eval freshness, repo hygiene, Paper-WorkFlow gates, stdlib unit tests, eval-harness lint + graded smoke, and the strict numeric benchmark (reference-gold drift fails the build). - External links and OpenSSF Scorecard run on their scheduled workflows.
当前收录版本
0.0.1
质量评测
满分 100。前四项来自 DeepSkill 收录的质量报告。
评分依据
提供LLM评估与基准测试的完整方法论体系
评分依据
触发词精准完整,覆盖LLM评估相关多种场景
评分依据
文档结构清晰,章节划分合理
评分依据
提供的Python代码示例逻辑完整,参数明确
计算方式
取前四项原始分值的算术平均,四舍五入取整。可通过收录报告核对各项评分与说明,运行记录需单独查证。
权限与使用前检查
根据当前主文档自动整理,命中内容可能是示例或防护说明;不是实际权限声明,也未覆盖附件、外部依赖或运行行为。
未从主文档识别到上述常见操作,不能据此判断无需权限或没有风险。
安全评级与访问范围
risk · 静态证据核验,不代表运行安全保证。
使用前请由宿主核对并授权实际文件读写、外部服务和联网目的地;本站的核验不会授予任何运行权限。
用户讨论
还没有已核实的点评。