mirror of https://github.com/razor-ai/soup.git
25 lines
598 B
YAML
25 lines
598 B
YAML
# RLHF Step 2: Train Reward Model
|
|
# Learn to score responses: is this output good or bad?
|
|
# Uses preference pairs (chosen vs rejected)
|
|
|
|
model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
|
|
data:
|
|
path: examples/data/chat_preferences.jsonl # must have chosen/rejected
|
|
format: sharegpt
|
|
task: reward_model
|
|
backend: transformers
|
|
quantization: null
|
|
batch_size: 8
|
|
gradient_accumulation_steps: 2
|
|
num_epochs: 2
|
|
learning_rate: 1e-4
|
|
lr_scheduler_type: cosine
|
|
warmup_ratio: 0.1
|
|
weight_decay: 0.01
|
|
max_seq_length: 2048
|
|
output_dir: ./output_reward_model/
|
|
seed: 42
|
|
logging_steps: 10
|
|
save_steps: 50
|
|
save_strategy: steps
|