Llama-3-8b-rm-700k

Llama3-8b Reward Model

This is the Llama3-8b-based Reward Model, trained using OpenRLHF, an efficient RLHF framework presented in the paper REINFORCE++: An Efficient RLHF Algorithm with Robustness to Both Prompt and Reward Models.

The model was trained with a combination of datasets available at OpenLLMAI/preference_700K.

Base SFT model: OpenRLHF/Llama-3-8b-sft-mixture

Training Configuration

Cosine Scheduler
Learning Rate: 9e-6
Warmup Ratio: 0.03
Batch Size: 256
Epoch: 1

Usage

You can use this model with the Hugging Face transformers library to score the quality of a generated response to a given prompt. The input format should match what the model was trained on (e.g., a full conversation turn using the Llama 3 chat template).

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "OpenRLHF/Llama-3-8b-rm-mixture" # This model ID
tokenizer = AutoTokenizer.from_pretrained(model_id)
# Ensure to load with appropriate torch_dtype, e.g., torch.bfloat16 for Llama models
model = AutoModelForSequenceClassification.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

# Example: Score responses to a user prompt
prompt = "Write a short poem about a cat."
response_good = "A feline friend, soft and sleek,\
Curled up warm, a purring peek.\
Through sunlit naps and playful chase,\
Graceful paws in every space."
response_bad = "Cats are okay. They sit sometimes. Dog is better."

# Apply the chat template for the full conversation turn (user prompt + assistant response)
# The `apply_chat_template` method structures the input as expected by the model.
messages_good = [
    {"role": "user", "content": prompt},
    {"role": "assistant", "content": response_good},
]
messages_bad = [
    {"role": "user", "content": prompt},
    {"role": "assistant", "content": response_bad},
]

input_ids_good = tokenizer.apply_chat_template(messages_good, return_tensors="pt", add_generation_prompt=False).to(model.device)
input_ids_bad = tokenizer.apply_chat_template(messages_bad, return_tensors="pt", add_generation_prompt=False).to(model.device)

# Get scores
with torch.no_grad():
    score_good = model(input_ids_good).logits.item()
    score_bad = model(input_ids_bad).logits.item()

print(f"Score for good response: {score_good:.2f}")
print(f"Score for bad response: {score_bad:.2f}")