mkd-hossain/keural-sft-18k

Keural-SFT-14.83B

Keural is a bilingual Korean–English Mixture-of-Experts language model trained from scratch. This is the SFT (Supervised Fine-Tuning) checkpoint at step 18,000, fine-tuned from the Keural stage-2 pretrained base using the ChatML instruction format.

Model Details

Property	Value
Architecture	Mixtral-style MoE (8 experts, top-2)
Parameters	14.83B total / ~7.42B active per token
Layers	24
Hidden size	4096
Attention heads	32 (GQA — 8 KV heads)
Expert intermediate size	5632
Context length	4096 tokens
Vocabulary	131,074 (131,072 SPM + `<
RoPE theta	500,000
Sliding window	512 (every other layer)
Dtype	bfloat16
Languages	Korean, English

Training Pipeline

Stage	Steps	Data
Pretraining Stage 1	100,000	Korean + English web corpus
Pretraining Stage 2	20,000	Korean + English web corpus (continued)
SFT (this checkpoint)	18,000	mkd-chanwoo/keural-SFT (1.14M ChatML samples)

SFT hyperparameters: LR 1e-5 → 1e-6 cosine, batch 64 effective (4 × 8 accum × 2 GPUs), max_seq 4096, weight_decay 0.05, 2× H200 SXM with FSDP FULL_SHARD.

Chat Format (ChatML)

This model uses ChatML format. You must use this exact format for good results.

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
안녕하세요! 오늘 날씨가 어때요?<|im_end|>
<|im_start|>assistant

The model generates until it produces <|im_end|> (token ID 131073).

How to Use

With transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "mkd-hossain/keural-sft-18k"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a helpful bilingual Korean-English assistant."},
    {"role": "user",   "content": "파이썬에서 리스트를 정렬하는 방법을 알려주세요."},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        repetition_penalty=1.1,
        do_sample=True,
        eos_token_id=131073,   # <|im_end|>
    )

response = tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=False)
response = response.split("<|im_end|>")[0].strip()
print(response)

With vLLM (recommended for serving)

pip install vllm

python -m vllm.entrypoints.openai.api_server \
    --model mkd-hossain/keural-sft-18k \
    --tokenizer mkd-hossain/keural-sft-18k \
    --dtype bfloat16 \
    --max-model-len 4096 \
    --tensor-parallel-size 1

Then call the OpenAI-compatible endpoint:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

response = client.chat.completions.create(
    model="mkd-hossain/keural-sft-18k",
    messages=[
        {"role": "system", "content": "You are a helpful bilingual assistant."},
        {"role": "user",   "content": "한국의 수도는 어디인가요?"},
    ],
    max_tokens=512,
    temperature=0.7,
)
print(response.choices[0].message.content)

With vLLM on multiple GPUs

python -m vllm.entrypoints.openai.api_server \
    --model mkd-hossain/keural-sft-18k \
    --dtype bfloat16 \
    --max-model-len 4096 \
    --tensor-parallel-size 2

Manual ChatML prompt (without apply_chat_template)

prompt = (
    "<|im_start|>system\n"
    "You are a helpful assistant.<|im_end|>\n"
    "<|im_start|>user\n"
    "Tell me about Seoul.<|im_end|>\n"
    "<|im_start|>assistant\n"
)

Special Tokens

Token	ID	Purpose
`<	im_start	>`
`<	im_end	>`
<bos>	1	Beginning of sequence
<eos>	2	End of sequence
<pad>	0	Padding

Important: Always set eos_token_id=131073 (<|im_end|>) when generating. If you use eos_token_id=2 (<eos>), generation may not stop correctly.

Recommended Generation Settings

generation_config = {
    "max_new_tokens": 512,
    "temperature": 0.7,
    "top_p": 0.9,
    "top_k": 50,
    "repetition_penalty": 1.1,
    "do_sample": True,
    "eos_token_id": 131073,
}

For factual / deterministic tasks use temperature=0.1, do_sample=False.

Limitations

• SFT training loss plateaued at ~1.96 (comparable models reach ~1.3–1.6). The model follows instructions but may produce repetitive or off-topic responses on complex prompts.
• The pretraining corpus contains Korean web data which skews the style toward informal language.
• Maximum context is 4096 tokens. Inputs longer than this will be truncated.
• This is an intermediate checkpoint — a DPO-aligned version will be released separately.

License

Apache 2.0