README.md

---
license: apache-2.0
language:
- en
library_name: transformers
tags:
- Tulu3
- Smollm
- SLMs
- Small
- Huggingface
- Allenai
- Reward Model
- RLVR
- RM
- Reward
base_model:
- SultanR/SmolTulu-1.7b-Instruct
datasets:
- allenai/llama-3.1-tulu-3-8b-preference-mixture
pipeline_tag: text-classification
---

# SmolLM2 1.7b Reward Model for RLVR Through Tulu 3!

![SmolTulu Banner](smoltulubanner.png)

SmolTulu-1.7b-RM is the reward model used to initialize the value function for [SmolTulu-1.7b-Reinforced](https://huggingface.co/SultanR/SmolTulu-1.7b-Reinforced), which leverages [AllenAI's Tulu 3 post-training pipeline](https://arxiv.org/abs/2411.15124) for reinforcement learning with verifiable rewards (RLVR). This model was trained using the same preference datasets and methodology as outlined in the Tulu 3 paper, adapted for the smaller model size.

## Evaluation

Evaluation results comparing SmolTulu-1.7b-RM against the Tulu 3 8b reward model on standard reward model benchmarks:

| Metric | SmolTulu-1.7b-RM | Tulu 3 8b RM |
|:-----------|:----------------:|:-------------:|
| RB Chat | *94.13* | **96.27** |
| RB Chat Hard | 43.64 | **55.92** |
| RB Safety | *75.54* | **84.05** |
| RB Reasoning | *68.01* | **76.50** |
| RB Average | *72.43* | **81.34** |
| UFB | *73.17* | **77.34** |

While the 1.7B reward model shows lower performance compared to the larger 8B model as expected, it still demonstrates strong capabilities across different evaluation categories, particularly in chat quality assessment.

## Usage

The reward model can be used with the transformers library:

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer

checkpoint = "SultanR/SmolTulu-1.7b-RM"
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint).to(device)

# Example of computing reward for a completion
def get_reward(prompt, completion):
    inputs = tokenizer(prompt + completion, return_tensors="pt").to(device)
    reward = model(**inputs).logits[0].item()
    return reward
```

## Training Details

The reward model was trained with the following settings:
- Base model: SmolTulu-1.7b-Instruct 
- Mixed precision: bfloat16
- Learning rate: 4e-5
- Effective batch size: 4
- Maximum sequence length: 2048 tokens
- Maximum prompt length: 2048 tokens
- Training epochs: 1  
- Training data: Tulu 3 8B preference mixture
- Evaluation data: UltraFeedback (cleaned)
- Gradient checkpointing enabled
- DeepSpeed Zero-3 for memory optimization

## Citation

```
@misc{alrashed2024smoltuluhigherlearningrate,
      title={SmolTulu: Higher Learning Rate to Batch Size Ratios Can Lead to Better Reasoning in SLMs}, 
      author={Sultan Alrashed},
      year={2024},
      eprint={2412.08347},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.08347}, 
}
```

The training methodology follows the Tulu 3 paper:

```
@article{lambert2024tulu3,
  title={TÜLU 3: Pushing Frontiers in Open Language Model Post-Training},
  author={Lambert, Nathan and Morrison, Jacob and Pyatkin, Valentina and others},
  year={2024},
  journal={arXiv preprint arXiv:2411.15124}
}
```