---
library_name: transformers
license: other
base_model: meta-llama/Llama-3.2-1B-Instruct
tags:
- llama-factory
- full
- generated_from_trainer
model-index:
- name: ScienceLLaMA-1B
  results: []
---

# ScienceLLaMA-3B
<p align="center">
• 🤗 <a href="https://huggingface.co/datasets/JingyaoLi/Science-Logits-1.2M" target="_blank">Data </a> 
• 🤗 <a href="https://huggingface.co/JingyaoLi/ScienceLLaMA-3b" target="_blank">ScienceLLaMA-3B </a> 
• 🤗 <a href="https://huggingface.co/JingyaoLi/ScienceLLaMA-1b" target="_blank">ScienceLLaMA-1B </a> 
• 🐱 <a href="https://github.com/dvlab-research/Logits-Based-Finetuning" target="_blank">Code</a> 
• 📃 <a href="https://arxiv.org/abs/2505.24461" target="_blank">Paper</a>
</p>

This model is a fine-tuned with **Logits-Based Finetuning** on the [JingyaoLi/Science-Logits-1.2M](https://huggingface.co/datasets/JingyaoLi/Science-Logits-1.2M), which integrates the strengths of supervised learning and knowledge distillation by combining teacher logits with ground truth labels. This preserves both correctness and linguistic diversity.

<div style="text-align: center;">
    <img src="./images/example.png" alt="example" />
</div>

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-06
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 4
- total_train_batch_size: 256
- total_eval_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 1

### Training results
<div style="text-align: center;">
    <img src="./images/performance.png" alt="performance" />
</div>


### Framework versions

- Transformers 4.45.0
- Pytorch 2.4.0+cu121
- Datasets 2.21.0
- Tokenizers 0.20.1