File size: 2,035 Bytes
c7ebaa1 2145d80 c7ebaa1 2145d80 c7ebaa1 bff2f94 2145d80 bff2f94 2145d80 bff2f94 c7ebaa1 2145d80 c7ebaa1 2145d80 c7ebaa1 | 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 | """
BioRLHF: Biological Reinforcement Learning from Human Feedback
A framework for fine-tuning LLMs on biological reasoning tasks using SFT, DPO,
and GRPO with verifier-based reward models for factual accuracy, calibrated
uncertainty, and chain-of-thought reasoning.
"""
__version__ = "0.2.0"
__author__ = "JangKeun Kim"
__email__ = "jangkeun.kim@med.cornell.edu"
def __getattr__(name):
"""Lazy imports for torch-dependent modules."""
if name == "SFTTrainingConfig":
from biorlhf.training.sft import SFTTrainingConfig
return SFTTrainingConfig
elif name == "run_sft_training":
from biorlhf.training.sft import run_sft_training
return run_sft_training
elif name == "DPOTrainingConfig":
from biorlhf.training.dpo import DPOTrainingConfig
return DPOTrainingConfig
elif name == "run_dpo_training":
from biorlhf.training.dpo import run_dpo_training
return run_dpo_training
elif name == "GRPOConfig":
from biorlhf.training.grpo import GRPOConfig
return GRPOConfig
elif name == "run_grpo_training":
from biorlhf.training.grpo import run_grpo_training
return run_grpo_training
elif name == "create_sft_dataset":
from biorlhf.data.dataset import create_sft_dataset
return create_sft_dataset
elif name == "load_dataset":
from biorlhf.data.dataset import load_dataset
return load_dataset
elif name == "evaluate_model":
from biorlhf.evaluation.evaluate import evaluate_model
return evaluate_model
elif name == "RewardComposer":
from biorlhf.verifiers.composer import RewardComposer
return RewardComposer
raise AttributeError(f"module 'biorlhf' has no attribute {name!r}")
__all__ = [
"__version__",
"SFTTrainingConfig",
"run_sft_training",
"DPOTrainingConfig",
"run_dpo_training",
"GRPOConfig",
"run_grpo_training",
"create_sft_dataset",
"load_dataset",
"evaluate_model",
"RewardComposer",
]
|