Spaces:
Running
Running
File size: 13,792 Bytes
ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 33bb385 ac627d5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 | """
Skill Invocation Environment Implementation.
Trains LLMs to decide WHEN to invoke procedural knowledge (skills) during
task-solving. Context cost model: each loaded skill costs context budget.
Reward has two distinct cost signals:
- Context hygiene (bloat_penalty): penalizes irrelevant skills still loaded at
submit time (-0.15 per skill).
- Token efficiency (token_waste_penalty): penalizes skills that were ever loaded
but turned out to be irrelevant, even if unloaded before submission (-0.05 per
skill). This captures cumulative token waste across the episode.
Actions: list, load, unload, submit (plus "invoke" as backward-compat alias for load).
"""
import random
from typing import Optional
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
from models import SkillInvocationAction, SkillInvocationObservation, SkillInvocationState
from task_bank import TASK_BANK, SKILL_BANK
from task_generator import TaskGenerator
DEFAULT_CONTEXT_BUDGET = 5
class SkillInvocationEnvironment(Environment):
"""
RL environment for training skill invocation decisions.
Episodes:
1. reset() samples a task, assembles skill catalog (relevant + distractors)
2. Agent can list, load, and unload skills (within context budget)
3. Agent submits a solution
4. Reward = correctness + precision + recall - bloat - token_waste
"""
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(
self,
use_procedural: bool = False,
procedural_seed: int = 0,
context_budget: int = DEFAULT_CONTEXT_BUDGET,
):
super().__init__()
self._state = SkillInvocationState(episode_id=str(uuid4()), step_count=0)
self._current_task = None
self._catalog_skill_ids: list[str] = []
self._messages: list[str] = []
self._use_procedural = use_procedural
self._task_generator = TaskGenerator(seed=procedural_seed) if use_procedural else None
self._episode_skills: dict = {}
self._context_budget = context_budget
# Per-instance RNG to avoid mutating global random state (concurrency-safe)
self._rng = random.Random()
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
**kwargs,
) -> SkillInvocationObservation:
"""Sample a random task and assemble the skill catalog."""
# Use a local RNG instance to avoid mutating global random state.
# This is concurrency-safe: parallel rollouts won't clobber each other's seeds.
if seed is not None:
self._rng = random.Random(seed)
else:
self._rng = random.Random()
if self._use_procedural and self._task_generator:
gen_seed = seed if seed is not None else self._rng.randint(0, 2**31)
result = self._task_generator.generate_with_seed(gen_seed)
task = result["task"]
self._episode_skills = result["skills"]
else:
task = self._rng.choice(TASK_BANK)
self._episode_skills = SKILL_BANK
self._current_task = task
# Build catalog: relevant + distractor skills, shuffled
catalog_ids = list(task["relevant_skills"]) + list(task["distractor_skills"])
self._rng.shuffle(catalog_ids)
self._catalog_skill_ids = catalog_ids
# Build catalog descriptions (short only, no full content)
skill_catalog = []
for sid in catalog_ids:
skill = self._episode_skills[sid]
skill_catalog.append({
"id": sid,
"name": skill["name"],
"description": skill["short_description"],
})
# Initialize state
eid = episode_id or str(uuid4())
self._state = SkillInvocationState(
episode_id=eid,
step_count=0,
task_id=task["id"],
loaded_skills=[],
skills_ever_loaded=[],
skills_invoked=[],
difficulty=task["difficulty"],
done=False,
context_budget_total=self._context_budget,
remaining_invocations=self._context_budget,
)
self._messages = [f"Episode started. Task: {task['id']} ({task['difficulty']})"]
return self._make_observation(
skill_content=None,
reward=0.0,
done=False,
)
def step(
self,
action: SkillInvocationAction,
timeout_s: Optional[float] = None,
**kwargs,
) -> SkillInvocationObservation:
"""Process a list, load, unload, or submit action."""
self._state.step_count += 1
if self._state.done:
self._messages.append("Episode already done. Call reset().")
return self._make_observation(
skill_content=None,
verification_result="Episode already finished.",
reward=0.0,
done=True,
)
action_type = action.action_type
# Backward compat: "invoke" is an alias for "load"
if action_type == "invoke":
action_type = "load"
if action_type == "load":
return self._handle_load(action)
elif action_type == "unload":
return self._handle_unload(action)
elif action_type == "submit":
return self._handle_submit(action)
else:
self._messages.append(f"Unknown action_type: {action.action_type}")
return self._make_observation(
skill_content=None,
reward=0.0,
done=False,
)
def _handle_load(self, action: SkillInvocationAction) -> SkillInvocationObservation:
"""Load a skill into context."""
skill_id = action.skill_id
if not skill_id:
self._messages.append("load action requires skill_id")
return self._make_observation(skill_content=None, reward=0.0, done=False)
if skill_id not in self._episode_skills:
self._messages.append(f"Unknown skill_id: {skill_id}")
return self._make_observation(skill_content=None, reward=0.0, done=False)
if skill_id not in self._catalog_skill_ids:
self._messages.append(f"Skill {skill_id} not in current catalog.")
return self._make_observation(skill_content=None, reward=0.0, done=False)
# Already loaded — no-op, but still return content
if skill_id in self._state.loaded_skills:
full_content = self._episode_skills[skill_id]["full_content"]
self._messages.append(f"Skill {skill_id} already loaded.")
return self._make_observation(skill_content=full_content, reward=0.0, done=False)
# Check context budget
if len(self._state.loaded_skills) >= self._state.context_budget_total:
self._messages.append(
f"Context budget full ({self._state.context_budget_total} skills loaded). "
"Unload a skill first."
)
return self._make_observation(skill_content=None, reward=0.0, done=False)
# Load skill
self._state.loaded_skills.append(skill_id)
if skill_id not in self._state.skills_ever_loaded:
self._state.skills_ever_loaded.append(skill_id)
# Backward compat
self._state.skills_invoked = list(self._state.skills_ever_loaded)
self._state.remaining_invocations = (
self._state.context_budget_total - len(self._state.loaded_skills)
)
full_content = self._episode_skills[skill_id]["full_content"]
skill_name = self._episode_skills[skill_id]["name"]
self._messages.append(
f"Loaded skill '{skill_name}' ({skill_id}). "
f"Context: {len(self._state.loaded_skills)}/{self._state.context_budget_total}"
)
return self._make_observation(
skill_content=full_content,
reward=0.0,
done=False,
)
def _handle_unload(self, action: SkillInvocationAction) -> SkillInvocationObservation:
"""Unload a skill from context."""
skill_id = action.skill_id
if not skill_id:
self._messages.append("unload action requires skill_id")
return self._make_observation(skill_content=None, reward=0.0, done=False)
if skill_id not in self._state.loaded_skills:
self._messages.append(f"Skill {skill_id} is not currently loaded.")
return self._make_observation(skill_content=None, reward=0.0, done=False)
self._state.loaded_skills.remove(skill_id)
self._state.remaining_invocations = (
self._state.context_budget_total - len(self._state.loaded_skills)
)
skill_name = self._episode_skills[skill_id]["name"]
self._messages.append(
f"Unloaded skill '{skill_name}' ({skill_id}). "
f"Context: {len(self._state.loaded_skills)}/{self._state.context_budget_total}"
)
return self._make_observation(skill_content=None, reward=0.0, done=False)
def _handle_submit(self, action: SkillInvocationAction) -> SkillInvocationObservation:
"""Handle a solution submission.
Reward = correctness + precision + recall - bloat - token_waste.
Two distinct cost signals:
- bloat_penalty (-0.15 per skill): penalizes irrelevant skills still
loaded at submit time (context hygiene).
- token_waste_penalty (-0.05 per skill): penalizes skills that were ever
loaded but turned out irrelevant, capturing cumulative token waste
across the episode (token efficiency).
"""
answer = action.answer or ""
task = self._current_task
# Run deterministic verifier
try:
task_correct = task["verifier"](answer)
except Exception:
task_correct = False
# Compute reward
loaded = set(self._state.loaded_skills)
ever_loaded = set(self._state.skills_ever_loaded)
relevant = set(task["relevant_skills"])
# 1. Correctness: +0.6
correctness = 0.6 if task_correct else 0.0
# 2. Precision: what fraction of loaded skills are relevant?
if len(loaded) > 0:
precision = len(loaded & relevant) / len(loaded)
else:
precision = 0.0
precision_bonus = 0.3 * precision
# 3. Recall: did you load all relevant skills?
if len(relevant) > 0:
recall = len(loaded & relevant) / len(relevant)
else:
recall = 1.0
recall_bonus = 0.1 * recall
# 4. Bloat: penalty for unnecessary skills loaded at submit time
unnecessary = loaded - relevant
bloat_penalty = -0.15 * len(unnecessary)
# 5. Token waste: penalty for skills ever loaded that were irrelevant
wasted = ever_loaded - relevant
token_waste_penalty = -0.05 * len(wasted)
total_reward = correctness + precision_bonus + recall_bonus + bloat_penalty + token_waste_penalty
total_reward = max(total_reward, -1.0)
self._state.done = True
verification_msg = (
f"{'CORRECT' if task_correct else 'INCORRECT'}. "
f"Reward: correctness={correctness:.2f}, "
f"precision={precision_bonus:.2f}, recall={recall_bonus:.2f}, "
f"bloat={bloat_penalty:.2f}, token_waste={token_waste_penalty:.2f}, "
f"total={total_reward:.2f}"
)
self._messages.append(f"Submitted answer. {verification_msg}")
return self._make_observation(
skill_content=None,
verification_result=verification_msg,
reward=total_reward,
done=True,
)
def _make_observation(
self,
skill_content: Optional[str],
reward: float,
done: bool,
verification_result: Optional[str] = None,
) -> SkillInvocationObservation:
"""Build an observation from current state."""
task = self._current_task
catalog = []
if task:
for sid in self._catalog_skill_ids:
if sid in self._episode_skills:
skill = self._episode_skills[sid]
catalog.append({
"id": sid,
"name": skill["name"],
"description": skill["short_description"],
})
# Build loaded skill contents
loaded_contents = {}
for sid in self._state.loaded_skills:
if sid in self._episode_skills:
loaded_contents[sid] = self._episode_skills[sid]["full_content"]
return SkillInvocationObservation(
task_description=task["description"] if task else "",
skill_catalog=catalog,
difficulty=self._state.difficulty,
loaded_skills=list(self._state.loaded_skills),
loaded_skill_contents=loaded_contents,
context_budget_used=len(self._state.loaded_skills),
context_budget_total=self._state.context_budget_total,
skill_content=skill_content,
remaining_invocations=(
self._state.context_budget_total - len(self._state.loaded_skills)
),
verification_result=verification_result,
skills_invoked=list(self._state.skills_ever_loaded),
messages=list(self._messages),
done=done,
reward=reward,
)
@property
def state(self) -> SkillInvocationState:
"""Get current episode state."""
return self._state
|