|
--- |
|
library_name: transformers |
|
license: cc-by-nc-4.0 |
|
language: |
|
- en |
|
- zh |
|
base_model: |
|
- meta-llama/Llama-3.2-1B-Instruct |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
# Kyara: Knowledge Yielding Adaptive Retrieval Augmentation for LLM Fine-tuning |
|
|
|
[](https://zenodo.org/badge/latestdoi/844304447) |
|
|
|
<p align="left"> |
|
π€ <a href="https://huggingface.co/zake7749/Llama-3.2-1B-it-chinese-kyara/">Hugging Face</a> ο½ π<a href="https://github.com/zake7749/kyara">Github</a> ο½ π <a href="#">Paper</a> ο½ π <a href="https://github.com/zake7749/kyara/blob/main/document/README_EN.md">English</a> ο½ π <a href="https://github.com/zake7749/kyara">Chinese</a> ο½ π» <a href="https://www.kaggle.com/code/zake7749/kyara-a-compact-yet-powerful-chinese-llm">Kaggle Notebook</a> |
|
</p> |
|
|
|
<div style="text-align: center;"> |
|
<img src="https://i.imgur.com/QiWlcYJ.jpeg" alt="kyara"/> |
|
</div> |
|
|
|
Kyara (Knowledge Yielding Adaptive Retrieval Augmentation) is an experimental project aimed at improving language models through knowledge retrieval processes. The project seeks to enhance the modelβs ability to adapt knowledge and improve language comprehension, particularly in underrepresented languages like Traditional Chinese. Given the relatively scarce availability of Traditional Chinese data compared to the vast corpus of English data used for model training, Kyara addresses this gap by expanding the limited corpus for this language. |
|
|
|
This is a preview model, with the stable version set to be released soon. |
|
|
|
## Benchmark |
|
|
|
All evaluations are conducted in a zero-shot setting. |
|
|
|
| Metric | Kyara-1b-it | Llama3.2-1b-it | |
|
|--------------------------|----------|-------------| |
|
| **[TMMLUPlus](https://huggingface.co/datasets/ikala/tmmluplus)** | **31.92** | 30.48 | |
|
|  - STEM | **32.56** | 29.74 | |
|
|  - Humanities | **30.60** | 29.89 | |
|
|  - Other | **31.08** | 30.32 | |
|
|  - Social-Science | **33.42** | 31.98 | |
|
| **[MMLU-Redux](https://github.com/yuchenlin/ZeroEval)** | **41.40** | 19.62βΊ | |
|
| **[GSM8K](https://github.com/yuchenlin/ZeroEval)** | 31.31 | **31.61** | |
|
| **[MATH-L5](https://github.com/yuchenlin/ZeroEval)** | **5.55** | 2.91 | |
|
| **[CRUX](https://github.com/yuchenlin/ZeroEval)** | **14** | 11 | |
|
| **[AlpacaEval](https://github.com/tatsu-lab/alpaca_eval)** | **10.79** | 7.39 | |
|
|
|
βΊ: Llama3.2-1b-it appears to have failed to follow the [output schema](https://github.com/WildEval/ZeroEval/blob/e3dd922cba9eeb8b76ed8212a81ee4cf6f30de2f/src/templates/MCQA.py) of ZeroEval on MMLU, with 45.28% of examples lacking answers, which has resulted in a lower MMLU score. |
