Spaces:
Sleeping
Sleeping
File size: 14,713 Bytes
ec4ae03 74e7a0b ec4ae03 74e7a0b ec4ae03 74e7a0b ec4ae03 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
AxiomForgeAI Math RL Environment.
Wraps CurriculumMathEnvironment from src/rl/math_environment_curriculum.py
to expose an OpenEnv-compatible interface (reset / step / state).
Episode semantics
-----------------
* reset() β Samples a new question from the adaptive curriculum (or a
grounded QA pair when a dataset is configured). Returns the
question in the observation; reward is 0.0.
* step(action) β Scores the agent's submitted solution with the full reward
pipeline (PRM + SymPy + format) and returns reward + feedback.
done=True always: one question per episode.
Environment variables
---------------------
AXIOMFORGE_DATA_PATH Path to a JSONL file with {"question", "gold_final"}
records (e.g. data/sft/gsm8k_sft.jsonl). When set,
the environment uses grounded QA pairs for questions
and ground-truth answer verification.
AXIOMFORGE_PRM_PATH HuggingFace model ID or local path for the Process
Reward Model (default: Qwen/Qwen2.5-Math-PRM-7B).
Set to "" to disable PRM scoring (uses SymPy only).
AXIOMFORGE_CURRICULUM_DIR
Directory where the CurriculumManager persists its
state between runs. Defaults to
"checkpoints/curriculum".
"""
from __future__ import annotations
import json
import logging
import os
import random
from pathlib import Path
from typing import Any, Dict, List, Optional
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
try:
from ..models import AxiomforgeaiAction, AxiomforgeaiObservation
except ImportError:
from models import AxiomforgeaiAction, AxiomforgeaiObservation
# ββ Heavy RL imports β fail gracefully so openenv validate passes even when
# the ML stack is not installed (e.g. lightweight CI / schema validation).
try:
import torch
from src.rl.math_environment_curriculum import CurriculumMathEnvironment
from src.rl.prm_scorer import ProcessRewardScorer
from src.sft.solution_format import extract_final_answer_numeric_str
_RL_AVAILABLE = True
except Exception as _rl_import_err: # pragma: no cover
torch = None # type: ignore[assignment]
_RL_AVAILABLE = False
CurriculumMathEnvironment = None # type: ignore[assignment,misc]
ProcessRewardScorer = None # type: ignore[assignment,misc]
extract_final_answer_numeric_str = None # type: ignore[assignment]
logger = logging.getLogger(__name__)
# Fallback question used during validation / when no dataset is configured.
_VALIDATION_QUESTION = (
"A store sells apples for $2 each and oranges for $3 each. "
"If Sarah buys 4 apples and 3 oranges, how much does she spend in total?"
)
_VALIDATION_GOLD = "17"
_VALIDATION_TOPIC = "basic_arithmetic"
_VALIDATION_DIFFICULTY = 0.1
def _load_qa_pairs(data_path: str) -> List[Dict[str, str]]:
"""Load {"question", "gold_final"} records from a JSONL file."""
pairs: List[Dict[str, str]] = []
p = Path(data_path)
if not p.exists():
logger.warning("AXIOMFORGE_DATA_PATH not found: %s", data_path)
return pairs
with p.open(encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
try:
rec = json.loads(line)
except json.JSONDecodeError:
continue
q = rec.get("question", "").strip()
g = rec.get("gold_final", "").strip()
if q and g:
pairs.append({"question": q, "gold_final": g})
logger.info("Loaded %d QA pairs from %s", len(pairs), data_path)
return pairs
class AxiomforgeaiEnvironment(Environment):
"""
AxiomForgeAI math RL environment for OpenEnv.
Uses CurriculumMathEnvironment from src/rl/ for adaptive question
selection and reward computation. When the ML stack is unavailable
(e.g. during schema validation), falls back to a lightweight mode
that uses only the installed openenv-core dependencies.
Supports concurrent WebSocket sessions β each client gets its own
instance with independent episode state.
"""
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(self) -> None:
self._state = State(episode_id=str(uuid4()), step_count=0)
# Per-episode state
self._current_question: str = ""
self._gold_final: str = ""
self._current_topic: str = ""
self._current_difficulty: float = 0.5
self._math_env: Optional[Any] = None # CurriculumMathEnvironment or None
if torch is not None:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
else:
device = "cpu"
if not _RL_AVAILABLE:
logger.warning(
"RL stack (torch/transformers/sympy) not available β "
"running in schema-validation mode with fixed fallback responses."
)
return
# ββ Load grounded QA pairs (optional) βββββββββββββββββββββββββββββ
grounded_qa_pairs: List[Dict[str, str]] = []
data_path = os.environ.get("AXIOMFORGE_DATA_PATH", "")
if data_path:
grounded_qa_pairs = _load_qa_pairs(data_path)
# ββ Load PRM scorer (optional) ββββββββββββββββββββββββββββββββββββ
prm: Optional[Any] = None # ProcessRewardScorer or None
prm_path = os.environ.get("AXIOMFORGE_PRM_PATH", "")
if prm_path:
try:
prm = ProcessRewardScorer(
model_name=prm_path,
device=device,
load_in_4bit=True,
)
logger.info("PRM loaded: %s", prm_path)
except Exception as exc:
logger.warning("PRM load failed (%s) β scoring uses SymPy only.", exc)
# ββ Create CurriculumMathEnvironment in scoring-only mode βββββββββ
# policy_model=None + tokenizer=None is safe when only reward-computation
# methods are called (compute_grounded_reward, sample_instruction).
# Generation methods (generate_with_logging, format_solution_prompt)
# are NOT called from the server step path β the agent supplies solutions.
curriculum_dir = os.environ.get(
"AXIOMFORGE_CURRICULUM_DIR", "checkpoints/curriculum"
)
try:
self._math_env = CurriculumMathEnvironment(
policy_model=None,
value_model=None,
tokenizer=None,
reference_questions=[qa["question"] for qa in grounded_qa_pairs],
grounded_qa_pairs=grounded_qa_pairs,
prm_scorer=prm,
curriculum_checkpoint_dir=curriculum_dir,
device=device,
)
logger.info(
"CurriculumMathEnvironment ready (scoring-only, %d QA pairs, PRM=%s)",
len(grounded_qa_pairs),
"yes" if prm else "no",
)
except Exception as exc:
logger.warning(
"CurriculumMathEnvironment init failed (%s) β "
"falling back to validation mode.",
exc,
)
self._math_env = None
# ------------------------------------------------------------------
# OpenEnv interface
# ------------------------------------------------------------------
def reset(
self,
qa: Optional[Dict[str, str]] = None,
) -> AxiomforgeaiObservation:
"""
Reset the environment and begin a new episode.
Args:
qa: Optional ``{"question": str, "gold_final": str}`` dict.
When supplied the environment is seeded with this specific
question and gold answer β used by the training loop for
difficulty-sampled grounded episodes. When omitted the
environment draws from its internal grounded QA pool (if
configured) or falls back to the curriculum instruction.
Returns:
AxiomforgeaiObservation with the question populated; reward=0.0.
"""
self._state = State(episode_id=str(uuid4()), step_count=0)
if qa is not None:
# Caller-supplied episode β honour it exactly.
self._current_question = qa.get("question", "").strip()
self._gold_final = qa.get("gold_final", "").strip()
self._current_topic = qa.get("topic", "grounded")
self._current_difficulty = float(qa.get("difficulty", 0.5))
elif self._math_env is not None:
try:
instruction, topic, difficulty = self._math_env.sample_instruction()
self._current_topic = topic
self._current_difficulty = float(difficulty)
if self._math_env.grounded_qa_pairs:
_qa = random.choice(self._math_env.grounded_qa_pairs)
self._current_question = _qa["question"]
self._gold_final = _qa["gold_final"]
else:
self._current_question = instruction
self._gold_final = ""
except Exception as exc:
logger.warning("sample_instruction failed, using fallback: %s", exc)
self._current_question = _VALIDATION_QUESTION
self._gold_final = _VALIDATION_GOLD
self._current_topic = _VALIDATION_TOPIC
self._current_difficulty = _VALIDATION_DIFFICULTY
else:
self._current_question = _VALIDATION_QUESTION
self._gold_final = _VALIDATION_GOLD
self._current_topic = _VALIDATION_TOPIC
self._current_difficulty = _VALIDATION_DIFFICULTY
return AxiomforgeaiObservation(
question=self._current_question,
topic=self._current_topic,
difficulty=self._current_difficulty,
feedback="",
done=False,
reward=0.0,
)
def step(self, action: AxiomforgeaiAction) -> AxiomforgeaiObservation: # type: ignore[override]
"""
Score the agent's submitted solution.
Uses compute_grounded_reward from CurriculumMathEnvironment when
available (PRM + SymPy + format scoring). Falls back to numeric
answer extraction when the full RL stack is not loaded.
Args:
action: AxiomforgeaiAction containing the solution text.
Returns:
AxiomforgeaiObservation with reward, feedback, and metadata.
done=True β one question per episode.
"""
self._state.step_count += 1
solution = action.solution
reward: float = 0.0
feedback: str = ""
metadata: Dict[str, Any] = {}
if self._math_env is not None and self._current_question:
try:
reward_result = self._math_env.compute_grounded_reward(
question=self._current_question,
solution=solution,
gold_final=self._gold_final,
)
reward = float(reward_result.get("combined_score", 0.0))
gt = reward_result.get("gt_match", False)
step_acc = reward_result.get("step_accuracy", 0.0)
lccp = reward_result.get("lccp", 0.0)
pred = reward_result.get("pred_final", "")
feedback = (
f"gt_match={gt} pred={pred!r} gold={self._gold_final!r} "
f"step_acc={step_acc:.2f} lccp={lccp:.2f}"
)
# Serialise reward breakdown into metadata; skip non-serialisable lists.
metadata = {
k: v
for k, v in reward_result.items()
if not isinstance(v, list)
}
except Exception as exc:
logger.warning("compute_grounded_reward failed: %s", exc)
reward, feedback, metadata = self._fallback_score(solution)
else:
reward, feedback, metadata = self._fallback_score(solution)
return AxiomforgeaiObservation(
question=self._current_question,
topic=self._current_topic,
difficulty=self._current_difficulty,
feedback=feedback,
done=True,
reward=reward,
metadata=metadata,
)
# ------------------------------------------------------------------
# Helpers
# ------------------------------------------------------------------
def _fallback_score(
self, solution: str
) -> tuple[float, str, Dict[str, Any]]:
"""Lightweight scoring used when the full RL stack is unavailable."""
pred: str = ""
if extract_final_answer_numeric_str is not None:
pred = extract_final_answer_numeric_str(solution) or ""
reward = 1.0 if pred and pred == self._gold_final else 0.0
feedback = f"pred={pred!r} gold={self._gold_final!r}"
return reward, feedback, {"pred_final": pred, "gold_final": self._gold_final}
def close(self) -> None:
"""
Persist curriculum state and release resources.
Call once at the end of a training run so the CurriculumManager's
per-topic statistics are saved to disk and can be resumed on the
next run. Safe to call multiple times.
"""
if self._math_env is not None:
try:
self._math_env.curriculum_manager.save_state(
iteration=self._math_env.curriculum_manager.current_iteration,
rollout=None,
)
logger.info(
"Curriculum state saved (iteration %d).",
self._math_env.curriculum_manager.current_iteration,
)
except Exception as exc:
logger.warning("close(): curriculum save failed β %s", exc)
@property
def state(self) -> State:
"""Return the current episode state (episode_id + step_count)."""
return self._state
|