菜单
Generate 1.5M realistic agent training trajectories from 495 real-world MCP servers without human annotation. Use to build tool-agentic training datasets with authentic tool execution, multi-turn conversations, and error handling.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill toucan-mcp-tool-agentic-dataset命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | toucan-mcp-tool-agentic-dataset |
| title | TOUCAN: Synthesizing Tool-Agentic Data from MCP Environments |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2510.01179 |
| keywords | ["MCP","agent-training","dataset-synthesis","tool-use","LLM-agents"] |
| description | Generate 1.5M realistic agent training trajectories from 495 real-world MCP servers without human annotation. Use to build tool-agentic training datasets with authentic tool execution, multi-turn conversations, and error handling. |
TOUCAN introduces the largest open-source tool-agentic dataset with 1.5M trajectories generated from real-world MCP (Model Context Protocol) servers. By leveraging authentic MCP execution rather than simulation, the dataset enables LLM agent training with realistic tool behavior, error patterns, and complex workflows.
Setup MCP server discovery and trajectory synthesis:
# Initialize TOUCAN dataset generation pipeline
from toucan import MCPDatasetSynthesizer, QualityFilter
synthesizer = MCPDatasetSynthesizer(
num_servers=495,
trajectory_budget=1_500_000,
parallel_workers=128,
server_timeout=30
)
# Discover available MCP servers
servers = synthesizer.discover_servers(
exclude_api_key_dependent=True, # avoid credential exposure
max_per_category=50
)
# Configure quality filtering (6-dimension validation)
quality_filter = QualityFilter(
dimensions={
"relevance": 0.85,
"correctness": 0.90,
"completeness": 0.80,
"diversity": 0.75,
"executability": 0.95,
"safety": 1.0
}
)
Generate trajectories with error handling and quality assurance:
# Generate realistic agent trajectories from MCP servers
trajectories = []
for server_config in servers:
# Stage 1: Instruction generation via LLM
instructions = synthesizer.generate_instructions(
server=server_config,
num_instructions_per_server=3000,
temperature=0.8
)
# Stage 2: Trajectory rollout with actual MCP execution
for instruction in instructions:
trajectory = synthesizer.rollout(
instruction=instruction,
server=server_config,
max_turns=15,
allow_parallel_calls=True,
error_handling="recovery"
)
# Stage 3: LLM-as-judge validation
score = quality_filter.evaluate(trajectory)
if score >= 0.80: # aggregated quality threshold
trajectories.append(trajectory)
# Stage 4: Safety screening
if not quality_filter.check_safety(trajectory):
continue
# Stage 5: Deduplication
if not synthesizer.is_duplicate(trajectory, existing=trajectories):
trajectories.append(trajectory)
Fine-tune agents on TOUCAN:
# Train LLM agents using TOUCAN trajectories
from transformers import AutoModelForCausalLM, Trainer, TrainingArguments
model = AutoModelForCausalLM.from_pretrained("Qwen2.5-32B")
training_args = TrainingArguments(
output_dir="./agent-model",
num_train_epochs=3,
per_device_train_batch_size=4,
learning_rate=5e-5,
max_seq_length=8192
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=toucan_dataset, # 1.5M trajectories
data_collator=agent_collator
)
trainer.train()
When to use TOUCAN:
When NOT to use:
Dataset statistics:
Training hyperparameters:
TOUCAN achieves state-of-the-art results on established benchmarks:
Relates to dataset synthesis for agent training and MCP ecosystem standardization.
基于 SOC 职业分类
