OnlySportsLM_196M

OnlySportsLM

Model Overview

OnlySportsLM is a 196M language model specifically designed and trained for sports-related natural language processing tasks. It is part of the larger OnlySports collection, which aims to advance domain-specific language modeling in sports.

Model Architecture

• Base architecture: RWKV-v6
• Parameters: 196 million
• Structure: 20 layers, 640 dimensions

Training

• Dataset: OnlySports Dataset (subset of 315B tokens out of 600B total)
• Training setup: 8 H100 GPUs
• Optimizer: AdamW
• Learning rate: Initially 6e-4, adjusted to 1e-4 due to observed loss spikes
• Context length: 1024 tokens

Performance

OnlySportsLM shows impressive performance on sports-related tasks:

• Outperforms previous SOTA 135M/360M models by 37.62%/34.08% on the OnlySports Benchmark
• Competitive with larger models like SomlLM 1.7B and Qwen 1.5B in the sports domain

<!-- For detailed performance metrics, please refer to our [technical report](https://github.com/chrischenhub/OnlySportsLM). -->

Usage

You can use this model for various sports-related content generation.

Download all files in this repo. Open RWKV_v6_demo.py for inference.

Limitations

• The model is specifically trained on sports-related content and may not perform as well on general topics
• Training was stopped at 315B tokens due to resource constraints, potentially limiting its full capabilities

Related Resources

• OnlySports Dataset
• Sports Text Classifier
• GitHub Repository

Citation

If you use OnlySportsLM in your research, please cite our paper.

Contact

For more information or inquiries about OnlySportsLM, please visit our GitHub repository.