Model Card: llm-course-hw2-reward-model

This model was trained as part of Homework 2 in the HSE LLM Course.
It serves as a Reward Model (RM) for PPO-style RLHF training.
The model learns to score pairs of (chosen, rejected) responses and assign higher reward to preferred outputs.

Model Description

• Base model: HuggingFaceTB/SmolLM-135M-Instruct
• Architecture: AutoModelForSequenceClassification
• Labels: 1 (scalar reward)
• Method: Reward modeling for RLHF
• Dataset: HumanLLMs/Human-Like-DPO-Dataset (converted into implicit preference format)
• Task: Score chosen responses higher than rejected ones

The model is used together with PPO policy/value models during alignment training.

Uses

This model is intended for educational purposes, including:

• demonstrating how to train reward models with TRL
• providing reward signals for PPO fine-tuning
• experimenting with RLHF pipelines

Not intended for production use.

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "sodeniZz/llm-course-hw2-reward-model"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

Training Hyperparameters

• Epochs: 1
• Batch size: 16 (per device)
• Max sequence length: 1024
• Learning rate: 3e-4
• Disable dropout: True
• Seed: 42
• Logging steps: 25
• Optimizer: AdamW