train_verifier.py

import torch
from datasets import load_dataset
from trl import RewardTrainer, RewardConfig
from peft import LoraConfig

HUB_ORG = 'narcolepticchicken'
MODEL = 'Qwen/Qwen3-4B'
DATASET = f'{HUB_ORG}/speculative-actions-verifier-pref'
OUTPUT = f'{HUB_ORG}/speculative-verifier-qwen3-4b'

print('Loading dataset...')
ds = load_dataset(DATASET)

print('Configuring LoRA...')
peft_config = LoraConfig(
    r=16, lora_alpha=32,
    target_modules=['q_proj', 'v_proj', 'k_proj', 'o_proj'],
)

print('Configuring Reward Training...')
config = RewardConfig(
    output_dir='/tmp/verifier-out',
    hub_model_id=OUTPUT,
    push_to_hub=True,
    learning_rate=2e-4,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    num_train_epochs=2,
    bf16=True,
    gradient_checkpointing=True,
    logging_strategy='steps',
    logging_steps=10,
    logging_first_step=True,
    disable_tqdm=True,
    report_to='trackio',
    run_name='verifier-reward-qwen3-4b',
)

print('Initializing Reward Trainer...')
trainer = RewardTrainer(
    model=MODEL,
    train_dataset=ds['train'],
    eval_dataset=ds['test'],
    args=config,
    peft_config=peft_config,
)

print('Training verifier...')
trainer.train()
trainer.push_to_hub()
print('Verifier training complete.')