textpolicy 0.0.1__py3-none-any.whl → 0.1.1__py3-none-any.whl

textpolicy/__init__.py

@@ -0,0 +1,53 @@
+"""
+TextPolicy: RL library for text generation with MLX.
+
+This module exposes the public API entry points for algorithms,
+training, generation, environment, and rewards.
+"""
+
+# Submodule imports for building the public API
+from . import algorithms, generation, training
+
+# Export RL algorithms as defined in textpolicy.algorithms.__all__
+from .algorithms import *  # noqa: F403,F401
+
+# Export text generation utilities (load_model, generate_tokens, etc.)
+from .generation import *  # noqa: F403,F401
+
+# Export training components (Trainer, RolloutManager, TrainingMetrics)
+from .training import *  # noqa: F403,F401
+
+# Export environment components and factory functions
+from .environment import (
+    TextGenerationEnvironment,
+    TextGenerationEnv,
+    create_text_generation_test_env,
+    validate_learning_progress,
+)
+
+# Export installation validation utilities
+from .validate import validate_installation
+
+# Export core reward functions and the reward decorator
+from .rewards.basic import length_reward, keyword_reward, perplexity_reward, accuracy_reward
+from .rewards.registry import reward, verifier
+
+# Build __all__ combining submodule __all__ lists and additional symbols
+__all__ = (
+    algorithms.__all__
+    + generation.__all__
+    + training.__all__
+    + [
+        "TextGenerationEnvironment",
+        "TextGenerationEnv",
+        "create_text_generation_test_env",
+        "validate_learning_progress",
+        "validate_installation",
+        "length_reward",
+        "keyword_reward",
+        "perplexity_reward",
+        "accuracy_reward",
+        "reward",
+        "verifier",
+    ]
+)

textpolicy/__main__.py

@@ -0,0 +1,8 @@
+"""Allow `python -m textpolicy` to invoke the CLI."""
+
+from .cli import main
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
+

textpolicy/algorithms/__init__.py

@@ -0,0 +1,54 @@
+# textpolicy/algorithms/__init__.py
+"""
+Reinforcement learning algorithms for MLX and Apple Silicon.
+
+GRPO: group-relative advantages with PPO-style clipping.
+GSPO: sequence-level importance sampling (sequence, token, and hybrid variants).
+"""
+
+from .grpo import (
+    compute_advantages,
+    compute_advantages_dr_grpo,
+    policy_loss,
+    grpo_loss,
+    compute_metrics,
+    entropy_bonus,
+    select_all_data,
+    select_recent_data
+)
+
+from .gspo import (
+    create_gspo_policy_loss,
+    create_gspo_metrics,
+    policy_loss_sequence,
+    policy_loss_hybrid,
+    create_policy_loss_hybrid,
+    policy_loss_token,
+    compute_metrics_sequence,
+    compute_metrics_hybrid,
+    compute_metrics_token,
+    select_gspo_data
+)
+
+__all__ = [
+    # GRPO functions
+    "compute_advantages",
+    "compute_advantages_dr_grpo",
+    "policy_loss",
+    "grpo_loss", 
+    "compute_metrics",
+    "entropy_bonus",
+    "select_all_data",
+    "select_recent_data",
+    # GSPO functions
+    "create_gspo_policy_loss",
+    "create_gspo_metrics",
+    "policy_loss_sequence",
+    "policy_loss_hybrid",
+    "create_policy_loss_hybrid",
+    "policy_loss_token",
+    "compute_metrics_sequence",
+    "compute_metrics_hybrid",
+    "compute_metrics_token",
+    "select_gspo_data"
+]