集成
将 Potato 与 AI 模型、众包平台连接,并导出到您常用的机器学习框架。
AI 与大语言模型集成
用 AI 辅助增强标注效率
Local LLMs (Ollama)
Run AI-assisted annotation with local LLMs using Ollama. Keep your data completely private.
查看文档 →HuggingFace
Access open-source models via HuggingFace Inference API for flexible AI assistance.
查看文档 →LangChain
Automatic trace ingestion from LangChain agents via callback handler. Capture full agent runs as annotation-ready traces.
查看文档 →OpenAI Vision
OpenAI vision models for multimodal annotation assistance on images and screenshots.
查看文档 →Hugging Face Hub
Publish a model trained on your annotations to a Hub repository, private by default, with a model card written from the training run.
查看文档 →Hugging Face Inference
Score an Inference Endpoint or serverless model on your labeled items, then show its predictions to annotators as pre-labels.
查看文档 →Roboflow
Send images to a hosted or self-hosted Roboflow model and review its boxes and outlines as pre-labels.
查看文档 →CVAT Nuclio functions
Call a CVAT serverless detector on your images and score its boxes against the boxes your annotators drew.
查看文档 →Label Studio ML backend
Ask a running Label Studio ML backend for predictions, score them on held-out items, and review them in Potato as pre-labels.
查看文档 →AI 驱动功能
- 智能标签建议
- 自动关键词高亮
- 质量检查辅助
- 预标注审核
- 解释生成
- 一致性检查
标注团队方案
使用您自己的团队或通过众包扩展规模
您自己的团队
推荐用于敏感数据在本地或您自己的服务器上运行 Potato,使用内部标注员。非常适合不能外泄的敏感数据、通过 IRB 审批的研究,或已有训练有素的标注团队的情况。
优势
或通过众包平台扩展规模
Prolific
Academic-friendly crowdsourcing with quality participants. Full integration with completion URLs and participant tracking.
功能特性
SONA Systems
University subject pools. Participants arrive with a SONA survey code, and Potato grants their credit from the server when they finish.
功能特性
Connect and other panels
Providers for CloudResearch Connect, Clickworker and Microworkers, and a generic provider for any panel that passes a participant ID.
功能特性
支持的数据格式
以任何常见格式导入数据
Text
.txt, .json, .jsonl
Images
.jpg, .png, .gif, .webp
Audio
.mp3, .wav, .ogg, .m4a
Video
.mp4, .webm, .mov
Documents
.pdf, .html
导出格式
将标注导出为主流机器学习格式
General
- JSONL
One JSON object per annotator per item
- CSV
Tabular export for spreadsheet analysis
- Parquet
Columnar format for pandas, DuckDB or Spark
NLP
- CoNLL
Standard format for NER and sequence labeling
- Hugging Face
Push annotations to the Hub as a dataset
- spaCy (via CoNLL)
Export CoNLL, then run spacy convert
Computer Vision
- COCO
MS COCO format for object detection
- YOLO
YOLO format for real-time detection
- Pascal VOC
XML bounding boxes, one file per image
Agent Trace Formats
Import agent traces from 15 frameworks for annotation. Convert via CLI or ingest in real time via webhook.
Agent Frameworks
- LangChain / LangSmith
Hierarchical runs, tool calls
- Langfuse
Polling or export; scores sent back
- OpenAI
Function calling, assistants
- Anthropic Claude
Tool use, thinking blocks
- MCP
Model Context Protocol sessions
- OpenTelemetry
Distributed span hierarchy
- ATIF
Standard interchange format
Web Agents
- WebArena
Screenshots, element targeting
- Web Agent
Mind2Web, Computer Use, raw recordings
Coding Agents
- Claude Code
Anthropic Messages API with tool_use
- Aider
Markdown chat with edit blocks
- SWE-Agent
Thought/action/observation trajectories
- SWE-bench
Model and gold patches, test lists
General
- ReAct
Generic thought/action/observation
- Multi-Agent
CrewAI, AutoGen, LangGraph
Agent Training Exports
Export agent annotations directly to training pipeline formats
命令行与 Python
用于自动化的编程接口
命令行
# Check a config for errors
potato validate config.yaml --strict
# Start the annotation server
potato start config.yaml
# Export annotations to COCO
python -m potato.export -c config.yaml \
-f coco -o exports/在 Python 中加载导出文件
# Export to CSV first:
# python -m potato.export -c config.yaml -f csv -o exports/
import pandas as pd
# One row per annotator per item
df = pd.read_csv("exports/annotations.csv")
print(df.groupby("user_id").size())