メニュー
Convert natural language to SQL for unknown database schemas by formulating the task as a partially observable MDP. Use dual-track GRPO (token-level masked advantages) to learn schema discovery and query generation jointly.
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 trust-sql-rl-text-to-sql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 | trust-sql-rl-text-to-sql |
| title | TRUST-SQL: Tool-Integrated Multi-Turn RL for Text-to-SQL over Unknown Schemas |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.16448 |
| keywords | ["Text-to-SQL","Reinforcement Learning","Schema Discovery","Agent Reasoning","Policy Gradient"] |
| description | Convert natural language to SQL for unknown database schemas by formulating the task as a partially observable MDP. Use dual-track GRPO (token-level masked advantages) to learn schema discovery and query generation jointly. |
Most text-to-SQL systems assume full schema information is available, but production databases contain hundreds of tables with noisy metadata. TRUST-SQL treats this realistically: an autonomous agent must discover relevant schema information iteratively before generating queries. By formulating the task as a Partially Observable Markov Decision Process (POMDP) and applying a novel dual-track reinforcement learning strategy (token-level masked advantages for credit assignment), the system achieves 30% absolute improvement on 4B models and 16% on 8B models while operating without pre-loaded schema.
The key innovation is token-level advantage masking that isolates exploration rewards from execution outcomes, solving the credit assignment problem in multi-turn tool use.
TRUST-SQL operates through an iterative four-phase schema discovery protocol:
The agent learns both schema discovery strategies (which tables/columns to explore) and query generation simultaneously through reinforcement learning.
Start by defining the POMDP state space and environment interface.
from dataclasses import dataclass
from typing import Dict, ,
json
:
natural_language_question:
discovered_tables: [, ]
explored_columns:
tool_history: []
current_turn:
() -> :
context =
context +=
table, info .discovered_tables.items():
context +=
col, col_type info.get(, {}).items():
context +=
context +=
context
:
():
.db = database_connection
.initial_tables_hidden = initial_tables_hidden
.all_tables = ._get_all_tables()
.all_columns = ._get_all_columns()
() -> :
tables = {}
query =
row .db.execute(query):
table_name = row[]
columns = {}
col_query =
col_row .db.execute(col_query):
columns[col_row[]] = col_row[]
tables[table_name] = {: columns}
tables
() -> []:
candidates = []
query_lower = query_text.lower()
table_name .all_tables.keys():
score = ._similarity_score(query_lower, table_name.lower())
candidates.append((table_name, score))
candidates.sort(key= x: x[], reverse=)
[table table, _ candidates[:limit]]
() -> []:
table_name .all_tables:
.all_tables[table_name]
() -> []:
table_name .all_tables:
[]
columns = .all_tables[table_name][]
candidates = []
col_name columns.keys():
score = ._similarity_score(query_text.lower(),
col_name.lower())
candidates.append((col_name, score))
candidates.sort(key= x: x[], reverse=)
[col col, _ candidates[:limit]]
() -> :
tokens1 = (text1.split())
tokens2 = (text2.split())
tokens1 tokens2:
intersection = (tokens1 & tokens2)
union = (tokens1 | tokens2)
intersection / union
() -> (, ):
:
result = .db.execute(, timeout=)
,
Exception e:
, (e)
Now implement the multi-turn agent with tool use and the dual-track reward system.
import torch
import torch.nn.functional as F
from torch.optim import AdamW
class SchemaDiscoveryAgent:
"""Multi-turn RL agent for iterative schema discovery."""
def __init__(self, language_model, environment, max_turns=10):
self.llm = language_model
self.env = environment
self.max_turns = max_turns
# Tool definitions
self.tools = {
'search_tables': self.env.search_tables,
'inspect_table': self.env.inspect_table,
'search_columns': self.env.search_columns,
'generate_sql': self._generate_sql_candidate
}
def step(self, state: PODMPState) -> (Dict, float):
"""One turn: decide next action and compute reward."""
context = state.to_context()
context += "\nAvailable tools: search_tables, inspect_table, search_columns, generate_sql\n"
context += "Choose next action:"
# LLM selects next action
action_text = self.llm.generate(context, max_tokens=100)
parsed_action = self._parse_action(action_text)
if not parsed_action:
return {'error': 'Could not parse action'}, -0.5
tool_name = parsed_action['tool']
tool_args = parsed_action['args']
# Execute tool
try:
result = self.tools[tool_name](**tool_args)
except Exception as e:
return {'error': str(e)}, -0.2
# Compute reward
reward = self._compute_reward(tool_name, result, state)
return {
'tool': tool_name,
'args': tool_args,
'result': result
}, reward
def _parse_action(self, action_text: str) -> Optional[Dict]:
"""Parse LLM output into structured action."""
# Simple parsing; in production use more robust extraction
if 'search_tables' in action_text:
query = action_text.split("query=")[-1].split('"')[1] if 'query=' in action_text else ''
return {'tool': 'search_tables', 'args': {'query_text': query}}
elif 'inspect_table' in action_text:
table = action_text.split('table=')[-1].split('"')[1] if 'table=' in action_text else ''
return {'tool': 'inspect_table', 'args': {'table_name': table}}
# ... other tools
return None
def _compute_reward(self, tool_name: str, result: Dict,
state: PODMPState) -> float:
"""Compute exploration vs execution reward."""
exploration_reward = 0.0
execution_reward = 0.0
if tool_name == 'search_tables':
# Reward discovering new tables
new_tables = len([t for t in result
if t not in state.discovered_tables])
exploration_reward = 0.1 * new_tables
elif tool_name == 'inspect_table':
# Reward discovering new columns
exploration_reward = 0.05 * len(result.get('columns', {}))
elif tool_name == 'generate_sql':
# Execution reward: correctness of generated query
is_valid, msg = self.env.validate_sql(result)
execution_reward = 0.5 if is_valid else -0.5
# Composite reward
return 0.6 * exploration_reward + 0.4 * execution_reward
def _generate_sql_candidate(self, discovered_schema: Dict,
question: str) -> str:
"""Generate SQL based on discovered schema."""
prompt = f"""Based on this schema:
{json.dumps(discovered_schema, indent=2)}
Answer the question with SQL: {question}"""
return self.llm.generate(prompt, max_tokens=200)
Finally, implement the dual-track GRPO with token-level masked advantages.
class TokenLevelMaskedGRPO:
"""Dual-track GRPO with token-level advantage masking for credit assignment."""
def __init__(self, model, agent, environment):
self.model = model
self.agent = agent
self.env = environment
self.optimizer = AdamW(model.parameters(), lr=1e-5)
def compute_masked_advantages(self, trajectory: List[Dict],
rewards: List[float],
action_types: List[str]) -> torch.Tensor:
"""Compute per-token advantages, masking by action type."""
advantages = []
for i, (action_type, reward) in enumerate(zip(action_types, rewards)):
# Separate advantage computation by tool type
if action_type == 'generate_sql':
# Execution reward applies to all tokens
advantage = reward
else:
# Exploration reward applies to decision tokens only
advantage = reward * 0.5
advantages.append(advantage)
return torch.tensor(advantages, dtype=torch.float32)
def update_step(self, question: str, max_turns: int = 10):
"""One RL update step."""
state = PODMPState(
natural_language_question=question,
discovered_tables={},
explored_columns=set(),
tool_history=[],
current_turn=0
)
trajectory = []
rewards = []
action_types = []
logprobs = []
# Rollout: collect trajectory
for turn in range(max_turns):
action_dict, reward = self.agent.step(state)
if 'error' in action_dict:
break
trajectory.append(action_dict)
rewards.append(reward)
action_types.append(action_dict['tool'])
# Compute log-probability for this action
context = state.to_context()
logprob = self._compute_action_logprob(context, action_dict)
logprobs.append(logprob)
# Update state
state.tool_history.append(action_dict)
state.current_turn += 1
if action_dict['tool'] == 'generate_sql':
break
# Compute masked advantages
advantages = self.compute_masked_advantages(trajectory, rewards,
action_types)
# Policy gradient loss with per-token masking
logprobs_tensor = torch.stack(logprobs)
loss = -(advantages * logprobs_tensor).mean()
# Update
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.optimizer.step()
return loss.item(), sum(rewards)
def _compute_action_logprob(self, context: str,
action_dict: Dict) -> torch.Tensor:
"""Compute log-probability of action under current policy."""
# Generate action from context, compute log-prob of actual action
logits = self.model(context)
# Simplified; in practice extract token probabilities
return torch.tensor(0.0) # Placeholder
def train(self, training_questions: List[str], num_steps: int = 100):
"""Train the agent."""
for step in range(num_steps):
question = training_questions[step % len(training_questions)]
loss, total_reward = self.update_step(question)
if (step + 1) % 10 == 0:
print(f"Step {step+1}: Loss={loss:.4f}, "
f"Total Reward={total_reward:.3f}")
Hyperparameters and When to Use:
When NOT to use:
Common Pitfalls: