|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- liswei/Taiwan-Text-Excellence-2B |
|
- liswei/PromptPair-TW |
|
- yentinglin/TaiwanChat |
|
base_model: apple/OpenELM |
|
language: |
|
- zh |
|
--- |
|
|
|
<center> |
|
<img src="https://huggingface.co/liswei/Taiwan-ELM/resolve/main/Taiwan%20ELM%20Logo.jpeg" alt="Efficient LLM for Taiwan"> |
|
</center> |
|
|
|
> Efficient LLM for Taiwan with open weights/datasets/checkpoints and affordable sizes (270M/1.1B) |
|
|
|
# Taiwan ELM |
|
|
|
Taiwan ELM is a family of Efficient LLMs for Taiwan base on [apple/OpenELM](https://huggingface.co/apple/OpenELM). |
|
The project aims to provide an efficient model for researchers without access to large-scale computing resources. |
|
|
|
The model is trained using [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) on 2B Traditional Chinese tokens and 500K instruction samples. |
|
We will extend the model to train on larger data sets and different base models if there is sufficient demand. |
|
|
|
## What is being released? |
|
|
|
We release both pre-trained **base models and instruction tuned variants** with 270M and 1.1B parameters. |
|
Along with the model, **datasets used to train the models** are also released. |
|
|
|
In an effort to improve transparency, training **checkpoints (including rng/optimizer state) and logs** are also released in the model page. |
|
|
|
List of released models: |
|
* [Taiwan-ELM-270M](https://huggingface.co/liswei/Taiwan-ELM-270M) |
|
* [Taiwan-ELM-1_1B](https://huggingface.co/liswei/Taiwan-ELM-1_1B) |
|
* [Taiwan-ELM-270M-Instruct](https://huggingface.co/liswei/Taiwan-ELM-270M-Instruct) |
|
* [Taiwan-ELM-1_1B-Instruct](https://huggingface.co/liswei/Taiwan-ELM-1_1B-Instruct) |
|
|
|
List of released datasets: |
|
* [liswei/Taiwan-Text-Excellence-2B](https://huggingface.co/datasets/liswei/Taiwan-Text-Excellence-2B) |
|
* [liswei/PromptPair-TW](https://huggingface.co/datasets/liswei/PromptPair-TW) |
|
* [liswei/wikinews-zhtw-dedup](https://huggingface.co/datasets/liswei/wikinews-zhtw-dedup) |
|
* [liswei/wikipedia-zhtw-dedup](https://huggingface.co/datasets/liswei/wikipedia-zhtw-dedup) |
|
* [liswei/coct-en-zhtw-dedup](https://huggingface.co/datasets/liswei/coct-en-zhtw-dedup) |
|
|
|
Some of the datasets are not used for training Taiwan ELM but also released: |
|
* [liswei/common-crawl-zhtw](https://huggingface.co/datasets/liswei/common-crawl-zhtw) |
|
* [liswei/c4-zhtw](https://huggingface.co/datasets/liswei/c4-zhtw) |
|
* [liswei/rm-static-zhTW](https://huggingface.co/datasets/liswei/rm-static-zhTW) |
|
|
|
## Usage Examples |
|
|
|
For instruction-tuned modesl, we adapt the [LLaMA2](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) template: |
|
```jinja2 |
|
<s>[INST] <<SYS>> |
|
{{ system_prompt }} |
|
<</SYS>> |
|
|
|
{{ user_message }} [/INST] |
|
``` |
|
|
|
The model could be load via `AutoModelForCausalLM` or `text-generation-inference` with `trust_remote_code=True`: |
|
```python |
|
taiwan_elm_270m = AutoModelForCausalLM.from_pretrained("liswei/Taiwan-ELM-270M", trust_remote_code=True) |
|
``` |
|
|
|
We also support additional generation methods and speculative generation, please find reference at [OpenELM#usage](https://huggingface.co/apple/OpenELM#usage). |
