|
|
--- |
|
|
base_model: google/gemma-3-270m-it |
|
|
library_name: transformers |
|
|
model_name: Router |
|
|
tags: |
|
|
- trl |
|
|
- sft |
|
|
- gemma3 |
|
|
licence: license |
|
|
datasets: |
|
|
- d-s-b/synthetic-reasoning-dataset |
|
|
--- |
|
|
|
|
|
# Model Card for Router |
|
|
|
|
|
This model is fine-tuned to serve as a router for reasoning tasks, classifying input queries into one of three categories: |
|
|
|
|
|
no_reasoning – Direct factual lookup or simple recall (e.g., "What is the capital of France?") |
|
|
|
|
|
low_reasoning – Requires light reasoning such as simple arithmetic, comparisons, or single logical steps (e.g., "If John has 5 apples and eats 2, how many are left?") |
|
|
|
|
|
high_reasoning – Requires multi-step reasoning, deep logical chains, or complex problem-solving (e.g., "Prove that the sum of two even numbers is always even"). |
|
|
|
|
|
## Quick start |
|
|
|
|
|
```python |
|
|
from transformers import pipeline |
|
|
|
|
|
pipe = pipeline("text-generation", model="d-s-b/Router") |
|
|
messages = [ |
|
|
{"role": "user", "content": "what is capital of india"} |
|
|
] |
|
|
pipe(messages) |
|
|
``` |
|
|
|
|
|
## Training Details |
|
|
|
|
|
Method: Supervised fine-tuning with SFTTrainer |
|
|
|
|
|
Objective: Multi-class classification with labels (no_reasoning, low_reasoning, high_reasoning) |
|
|
|
|
|
Dataset: Custom dataset of queries annotated with reasoning levels. |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
## Limitations & Bias |
|
|
|
|
|
May misclassify borderline queries (e.g., between low_reasoning and high_reasoning). |
|
|
|
|
|
Performance depends on the diversity of training data. |
|
|
|
|
|
Inherits any biases from the base Gemma 3 270M model. |
|
|
|
|
|
### Framework versions |
|
|
|
|
|
- TRL: 0.21.0 |
|
|
- Transformers: 4.55.1 |
|
|
- Pytorch: 2.6.0+cu124 |
|
|
- Datasets: 4.0.0 |
|
|
- Tokenizers: 0.21.4 |
|
|
|
|
|
## Citations |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
```bibtex |
|
|
@misc{vonwerra2022trl, |
|
|
title = {{TRL: Transformer Reinforcement Learning}}, |
|
|
author = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec}, |
|
|
year = 2020, |
|
|
journal = {GitHub repository}, |
|
|
publisher = {GitHub}, |
|
|
howpublished = {\url{https://github.com/huggingface/trl}} |
|
|
} |
|
|
@article{gemma_2025, |
|
|
title={Gemma 3}, |
|
|
url={https://arxiv.org/abs/2503.19786}, |
|
|
publisher={Google DeepMind}, |
|
|
author={Gemma Team}, |
|
|
year={2025} |
|
|
} |
|
|
|
|
|
|
|
|
``` |
