メニュー
Synthesize diverse, verifiable training tasks by executing real tools first, then reverse-deriving tasks from execution traces. Ensure diversity across tools and reasoning patterns while maintaining grounding by construction.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ADu2021/skillXiv --skill dive-diverse-task-synthesisコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
SOC 職業分類に基づく
SKILL.md を表示中
| name | dive-diverse-task-synthesis |
| title | DIVE: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.11076 |
| keywords | ["Task Synthesis","Tool Use","Evidence-Based","Agentic Training","Generalization"] |
| description | Synthesize diverse, verifiable training tasks by executing real tools first, then reverse-deriving tasks from execution traces. Ensure diversity across tools and reasoning patterns while maintaining grounding by construction. |
Traditional task synthesis generates queries first, then validates them—brittle and often fails. DIVE inverts this: it executes diverse tools in real environments to generate evidence, then reverse-derives tasks strictly entailed by these executions. This ensures tasks are executable by construction and verifiable from observable tool outputs.
The approach guarantees structural diversity (heterogeneous tool-use patterns) and validity without simulation artifacts.
DIVE operates through three phases:
This ensures every generated task is grounded in real tool behavior with verifiable reference answers.
Construct decoupled resource sets for breadth and coverage.
import json
from collections import defaultdict
class DiverseResourcePool:
def __init__(self):
self.tool_pool = {}
self.domain_concepts = defaultdict(list)
self.query_exemplars = []
def load_tool_pool(self, tools_file):
(tools_file, ) f:
tools = json.load(f)
tool_name, tool_spec tools.items():
.tool_pool[tool_name] = {
: tool_spec.get(),
: tool_spec.get(),
: tool_spec.get(),
: tool_spec.get(),
: ._get_tool_executor(tool_name)
}
():
():
executor
():
concept concepts:
.domain_concepts[domain].append({
: concept,
:
})
():
exemplar exemplars:
.query_exemplars.append({
: exemplar[],
: exemplar.get(, []),
: exemplar.get()
})
Run diverse tools to generate execution traces as ground truth.
class ToolExecutionEngine:
def __init__(self, resource_pool):
self.resource_pool = resource_pool
def execute_tool_diversity(self, num_executions=1000, domain_balance=True):
"""
Execute tools to generate diverse evidence traces.
domain_balance: True to sample evenly across domains
"""
execution_traces = []
domains = list(self.resource_pool.domain_concepts.keys())
domain_counts = {d: 0 for d in domains}
for exec_idx in range(num_executions):
# Select tool with balanced domain sampling
if domain_balance:
# Pick undersampled domain
domain = min(domains, key=lambda d: domain_counts[d])
else:
domain = random.choice(domains)
tools_in_domain = [
t for t, spec in self.resource_pool.tool_pool.items()
if spec['category'] == domain
]
if not tools_in_domain:
continue
# Select tool
tool_name = random.choice(tools_in_domain)
tool_spec = self.resource_pool.tool_pool[tool_name]
# Generate valid arguments for tool
tool_args = self._generate_valid_arguments(tool_spec)
# Execute tool
try:
result = tool_spec['executor'](tool_args)
trace = {
'tool': tool_name,
'arguments': tool_args,
'output': result,
'domain': domain,
'success': result is not None
}
execution_traces.append(trace)
domain_counts[domain] += 1
except Exception as e:
print(f"Execution failed for {tool_name}: {e}")
return execution_traces
def _generate_valid_arguments(self, tool_spec):
"""Generate arguments matching tool input schema."""
args = {}
for param_name, param_schema in tool_spec['inputs'].items():
if param_schema['type'] == 'string':
args[param_name] = self._sample_string_value(
param_schema.get('description')
)
elif param_schema['type'] == 'number':
args[param_name] = random.uniform(
param_schema.get('minimum', 0),
param_schema.get('maximum', 100)
)
return args
def _sample_string_value(self, description):
"""Sample realistic string based on parameter description."""
# Simplified: use description to infer domain
return f"sample_{hash(description) % 1000}"
Generate task descriptions that are strictly entailed by execution traces.
class TaskDeriver:
def __init__(self, llm_model):
self.llm = llm_model
def derive_task_from_trace(self, execution_trace):
"""
Generate task description from execution trace.
Ensures task is strictly entailed by observable outputs.
"""
tool_name = execution_trace['tool']
tool_args = execution_trace['arguments']
output = execution_trace['output']
# Prompt LLM to derive task
prompt = f"""Given a tool execution trace, derive a task description that is strictly entailed.
Tool: {tool_name}
Arguments: {json.dumps(tool_args)}
Output: {json.dumps(output)}
Generate a task description that:
1. Can be accomplished using {tool_name}
2. Is verifiable from the output
3. Matches the provided arguments as natural parameters
Task:"""
task_description = self.llm.generate(prompt, max_tokens=150)
return task_description.strip()
def synthesize_diverse_task_sequence(
self,
execution_traces,
num_tasks=100,
ensure_diversity=True
):
"""
Generate task-answer pairs from traces, with diversity guarantees.
"""
synthesized_tasks = []
if ensure_diversity:
# Group traces by tool to ensure coverage
traces_by_tool = defaultdict(list)
for trace in execution_traces:
traces_by_tool[trace['tool']].append(trace)
# Sample evenly from tools
tools = list(traces_by_tool.keys())
num_per_tool = num_tasks // len(tools)
for tool in tools:
tool_traces = traces_by_tool[tool]
sampled = random.sample(
tool_traces,
min(num_per_tool, len(tool_traces))
)
for trace in sampled:
task = self.derive_task_from_trace(trace)
synthesized_tasks.append({
'task_description': task,
'tool': trace['tool'],
'arguments': trace['arguments'],
'reference_answer': trace['output'],
'domain': trace['domain']
})
else:
# Sample without diversity guarantee
sampled_traces = random.sample(
execution_traces,
min(num_tasks, len(execution_traces))
)
for trace in sampled_traces:
task = self.derive_task_from_trace(trace)
synthesized_tasks.append({
'task_description': task,
'tool': trace['tool'],
'arguments': trace['arguments'],
'reference_answer': trace['output'],
'domain': trace['domain']
})
return synthesized_tasks
Validate that generated tasks are indeed entailed by evidence.
class TaskVerifier:
def __init__(self, llm_model):
self.llm = llm_model
def verify_task_answer_pair(self, task, reference_answer):
"""
Verify that reference answer satisfies task requirements.
"""
prompt = f"""Task: {task['task_description']}
Reference Answer: {json.dumps(reference_answer)}
Does the reference answer correctly accomplish the task? Answer 'Yes' or 'No'."""
response = self.llm.generate(prompt, max_tokens=10)
is_valid = 'yes' in response.lower()
return is_valid
def verify_task_batch(self, synthesized_tasks, sample_fraction=0.1):
"""
Spot-check synthesized tasks for quality.
"""
num_to_check = max(1, int(len(synthesized_tasks) * sample_fraction))
sampled = random.sample(synthesized_tasks, num_to_check)
valid_count = 0
for task in sampled:
is_valid = self.verify_task_answer_pair(task, task['reference_answer'])
if is_valid:
valid_count += 1
validity_rate = valid_count / len(sampled)
return {
'validity_rate': validity_rate,
'checked': len(sampled),
'valid': valid_count
}
Fine-tune agent on synthesized tasks, then apply RL.
def train_agent_with_div_tasks(
model,
synthesized_tasks,
num_sft_epochs=3,
num_rl_steps=10000
):
"""
Full training pipeline: SFT + RL with synthesized tasks.
"""
# Phase 1: Supervised Fine-Tuning
sft_optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
for epoch in range(num_sft_epochs):
total_loss = 0
for batch in batch_iter(synthesized_tasks, batch_size=32):
# Forward pass
outputs = model(batch['task_description'])
# Loss: cross-entropy with reference answers
loss = compute_task_loss(outputs, batch['reference_answer'])
sft_optimizer.zero_grad()
loss.backward()
sft_optimizer.step()
total_loss += loss.item()
print(f"SFT Epoch {epoch + 1}: Loss={total_loss / len(synthesized_tasks):.4f}")
# Phase 2: Reinforcement Learning
rl_optimizer = torch.optim.Adam(model.parameters(), lr=5e-6)
for rl_step in range(num_rl_steps):
# Sample task batch
batch = random.sample(synthesized_tasks, batch_size=32)
# Generate trajectories
trajectories = [model.sample_trajectory(t['task_description']) for t in batch]
# Evaluate using reference answers (verifiable)
rewards = [
1.0 if trajectory == t['reference_answer'] else 0.0
for trajectory, t in zip(trajectories, batch)
]
# GRPO-style RL update (simplified)
loss = compute_rl_loss(trajectories, rewards, model)
rl_optimizer.zero_grad()
loss.backward()
rl_optimizer.step()
if rl_step % 1000 == 0:
print(f"RL Step {rl_step}: Avg Reward={sum(rewards) / len(rewards):.2f}")
return model
When to Use:
When NOT to Use:
Hyperparameter Tuning:
Common Pitfalls: