seonglae/yokhal-md

Yokhal (욕쟁이 할머니)

<!-- Provide a quick summary of what the model is/does. --> Korean Chatbot based on Google Gemma

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

• —Fine-tuned by: Seonglae Cho
• —Model type: Gemma
• —Language(s) (NLP): Korean, English
• —Finetuned from model: Gemma-2b-it

Model Sources

<!-- Provide the basic links for the model. -->

• —Repository: https://github.com/seonglae/yokhal
• —Demo: https://huggingface.co/spaces/seonglae/yokhal

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

Korean Chatbot with Internet culture

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

Use the code below to get started with the model.

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16,
                                             device_map="auto" if device is None else device, 
                                             attn_implementation="flash_attention_2") # if flash enabled
sys_prompt = '한국어로 대답해'
texts = ['안녕', '서울은 오늘 어때']
chats = list(map(lambda t: [{'role': 'user', 'content': f'{sys_prompt}\n{t}'}], texts)) # ChatML format
prompts = list(map(lambda p: tokenizer.apply_chat_template(p, tokenize=False, add_generation_prompt=True), chats))
input_ids = tokenizer(prompts, return_tensors="pt", padding=True).to("cuda" if device is None else device)
outputs = model.generate(**input_ids, max_new_tokens=100, repetition_penalty=1.05)
for output in outputs:
  print(tokenizer.decode(output, skip_special_tokens=True), end='\n\n')

Training Details

Trained on 2 x RTX3090

More Information on Github source code

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

[More Information Needed]

Training Procedure

1. 1.Weight Initialized from Internet comments dataset
2. 2.Trained on Korean Namuwiki dataset until step 80000 (30000 step is on main branch because of repetition issue above there)
3. 3.seq_length 1024 with dataset packing
4. 4.batch 3 per device
5. 5.lr 1e-5
6. 6.optim adafactor
7. 7.Instruction tuning on Korean Instruction Dataset using QLoRa (not on main)
8. 8.seq_length 2048
9. 9.lr 2e-4

Preprocessing [optional]

Gemma do not support explicit system prompt in ChatML, so I trained putting system prompt before user message like below

if (chat[0]['role'] == 'system'):
  chat[1]['content'] = f"{chat[0]['content']}\n{chat[1]['content']}"
  chat = chat[1:]
try:
  prompt = tokenizer.apply_chat_template(chat, tokenize=False)

Source Code

Training Hyperparameters

• —Training regime: [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->

Speeds, Sizes, Times [optional]

<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->

[More Information Needed]

Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

Testing Data, Factors & Metrics

Testing Data

<!-- This should link to a Dataset Card if possible. -->

[More Information Needed]

Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->

[More Information Needed]

Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

[More Information Needed]

Results

[More Information Needed]

Summary