|
--- |
|
license: apache-2.0 |
|
language: |
|
- th |
|
- en |
|
datasets: |
|
- laion/OIG |
|
- databricks/databricks-dolly-15k |
|
- thaisum |
|
- scb_mt_enth_2020 |
|
- garage-bAInd/Open-Platypus |
|
- iapp_wiki_qa_squad |
|
- pythainlp/han-instruct-dataset-v1.0 |
|
- cognitivecomputations/dolphin |
|
- Hello-SimpleAI/HC3 |
|
- Muennighoff/xP3x |
|
--- |
|
# Model Card for WangChanLion 7B - The Multilingual Instruction-Following Model |
|
|
|
WangChanLion is a Multilingual, instruction-finetuned on Southeast Asian Languages SEA-LION 7B using open-source, commercially permissible datasets sample from LAION OIG chip2 and infill_dbpedia, DataBricks Dolly v2, OpenAI TL;DR, Hello-SimpleAI HC3, dolphin, iapp_wiki_qa_squad, thaisum, xlsum, scb_mt_enth_2020, han dataset, xp3x and Open-Platypus, a total of ~500k samples. Non-commercial datasets were filtered out. Released under apache 2.0 license. The models are trained to perform a subset of instruction-following tasks we found most relevant: reading comprehension, brainstorming, and creative writing. In this model, we focus on Thai and English datasets. We perform Vicuna-style evaluation using human evaluation. In a similar manner to Dolly v2, we only use open-source, commercially permissive pretrained models and datasets. Our models are neither restricted by non-commercial clauses like LLaMA-based models nor non-compete clauses like models that use self-instruct datasets from ChatGPT. |
|
|
|
- Developers: PyThaiNLP and VISTEC-depa AI Research Institute of Thailand |
|
- Model type: SEA-LION 7B (MPT architecture) |
|
|
|
## Model Sources |
|
- Repository: https://github.com/vistec-AI/WangchanLion |
|
- Demo: [demo_WangchanLion.ipynb - Colaboratory](https://colab.research.google.com/drive/1y_7oOU3ZJI0h4chUrXFL3K4kelW_OI2G?usp=sharing#scrollTo=4yN3Bo6iAH2L) |
|
|
|
# Direct Use |
|
Intended to be used as an instruction-following model for reading comprehension, brainstorming, and creative writing. |
|
|
|
# Downstream Use |
|
The model can be finetuned for any typical instruction-following use cases. |
|
|
|
# Out-of-Scope Use |
|
We do not expect the models to perform well in math problems, reasoning, and factfulness. |
|
|
|
# Bias, Risks, and Limitations |
|
We noticed similar limitations to other finetuned instruction followers, such as math problems, reasoning, and factfulness. Even though the models do not perform on the level that we expect them to be abused, they do contain undesirable biases and toxicity and should be further optimized for your particular use cases. |
|
|
|
# Recommendations |
|
Users (both direct and downstream) should be made aware of the risks, biases, and limitations of the model. More information is needed for further recommendations. |
|
|
|
# How to Get Started with the Model |
|
Use the code [here](https://colab.research.google.com/drive/1y_7oOU3ZJI0h4chUrXFL3K4kelW_OI2G?usp=sharing#scrollTo=4yN3Bo6iAH2L) to get started with the model. |
|
|
|
Or |
|
|
|
```python |
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
tokenizer = AutoTokenizer.from_pretrained( "airesearch/WangchanLion7B", trust_remote_code=True) |
|
model = AutoModelForCausalLM.from_pretrained( |
|
"airesearch/WangchanLion7B", trust_remote_code=True, |
|
return_dict=True, |
|
load_in_8bit=True , |
|
device_map="auto", |
|
torch_dtype=torch.float16, |
|
offload_folder="./", |
|
low_cpu_mem_usage=True, |
|
) |
|
def get_prompt(question: str,context: str = None) -> str: |
|
if context is not None: |
|
return """พื้นหลัง:\n\n{context}\n\nคำถาม:{question}\n\nตอบ:""".format(context=context, question=question) |
|
return """คำถาม:{question}\n\nตอบ:""".format(question=question) |
|
|
|
question = "เกิดอะไรขึ้นที่เทียนอันเหมินตอนปี 1989" |
|
full_prompt = get_prompt(question=question) |
|
tokens = tokenizer(full_prompt, return_tensors="pt").to("cuda") |
|
output = model.generate( |
|
input_ids=tokens['input_ids'], |
|
attention_mask=tokens['attention_mask'], |
|
max_new_tokens=256, |
|
early_stopping=True, |
|
top_k=50, top_p=0.95, |
|
do_sample=True, |
|
temperature=0.3, |
|
repetition_penalty = 1.2, |
|
eos_token_id = tokenizer.eos_token_id, |
|
) |
|
print(tokenizer.decode(output[0], skip_special_tokens=True)) |
|
``` |
|
|
|
# Training Details |
|
## Training Data |
|
Finetuning datasets are sourced from [LAION OIG chip2 and infill_dbpedia (Apache-2.0)](https://huggingface.co/datasets/laion/OIG), [DataBricks Dolly v2 (Apache-2.0)](https://github.com/databrickslabs/dolly), [OpenAI TL;DR (MIT)](https://github.com/openai/summarize-from-feedback), [Hello-SimpleAI HC3 (CC-BY SA)](https://huggingface.co/datasets/Hello-SimpleAI/HC3), [dolphin](https://huggingface.co/datasets/ehartford/dolphin), [iapp_wiki_qa_squad](https://huggingface.co/datasets/iapp_wiki_qa_squad) , [thaisum](https://huggingface.co/datasets/thaisum), [xlsum](https://huggingface.co/datasets/csebuetnlp/xlsum), [scb_mt_enth_2020](https://huggingface.co/datasets/scb_mt_enth_2020), [han dataset](https://huggingface.co/datasets/pythainlp/han-instruct-dataset-v1.0), [xp3x](https://huggingface.co/datasets/Muennighoff/xP3x) and [Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus). |
|
## Training regime |
|
- QLoRA with 4 A100 (40GB) |
|
|
|
|
|
# Evaluation |
|
We performed human and machine evaluations on XQuAD zero-shot and one-shot settings: |
|
## XQuAD |
|
| Model | Exact Match (Zero-shot) | F1 (Zero-shot) | Exact Match (One-shot) | F1 (One-shot) | |
|
|:--------------:|:-----------------------:|:--------------:|:----------------------:|:-------------:| |
|
| openthaigpt7B | 18.57 | 28.40 | 30.42 | 39.76 | |
|
| SeaLLM7B | 6.30 | 21.67 | 34.45 | 47.80 | |
|
| Typhoon-7b | - | 34.46 | - | **54.03** | |
|
| WangchanLion7B | **37.56** | **49.84** | **39.24** | 51.07 | |
|
|
|
## iAPP Wiki QA |
|
| Model | Exact Match (Zero-shot) | F1 (Zero-shot) | Exact Match (One-shot) | F1 (One-shot) | |
|
|:--------------:|:-----------------------:|:--------------:|:----------------------:|:-------------:| |
|
| openthaigpt7B | 22.06 | 40.07 | 31.39 | 47.98 | |
|
| SeaLLM7B | 8.25 | 34.40 | 40.05 | 58.27 | |
|
| WangchanLion7B | **55.48** | **67.93** | **56.43** | **68.85** | |
|
|
|
# What WangchanLion offers: |
|
- Transparent pretrained model: The development of SEA-LION is community-driven, with different ASEAN collaborators contributing pretraining datasets. The SEA-LION developers ensure that all datasets are safe and can be utilized without commercial restrictions. This transparency extends to the provision of pretraining code, ensuring anyone can replicate SEA-LION using the provided datasets. |
|
- Transparent finetuning data: In the spirit of open science, we make the finetuning data for WangchanLion accessible to all. This commitment to openness empowers the community by providing complete visibility into the instruction finetuning data that shapes WangchanLion. |
|
- Transparent finetuning code: The finetuning code for WangchanLion is readily available for distribution. By sharing our methods and processes, we invite others to learn from, build upon, and innovate alongside us. |