Korean 1B hybrid base model (Gated DeltaNet + GQA), trained from scratch

A 1.0B-parameter Korean language model pretrained from random initialisation on 25.17B tokens (89% Korean). Architecture, tokenizer, Gated DeltaNet Triton kernels, Muon optimizer and training loop are our own code (PyTorch + Triton).

Model

parameters 1.002B
layers 20, interleaved 3:1 ratio of Gated DeltaNet to attention layers
width d_model 2048, SwiGLU FFN 4352, RMSNorm (pre-norm)
attention layers GQA, 16 query / 4 KV heads Γ— 128, RoPE (ΞΈ=10000), QK-norm
Gated DeltaNet layers 16 heads, d_k = d_v = 128, short causal conv (k=4), L2-normalised q/k, output gate; chunkwise (C=64) WY-form kernel in Triton
vocabulary 48,000 (byte-level BPE trained on this corpus)
context 2048 tokens

Training corpus

source Hugging Face dataset tokenized size after filtering tokens seen in training
FineWeb-2, Korean (kor_Hang, train split) HuggingFaceFW/fineweb-2 38.28B (52.65M of 60.87M docs) 21.62B (β‰ˆ0.56 epochs)
Korean Wikipedia (2023-11-01 dump) wikimedia/wikipedia, 20231101.ko 0.28B (455k articles) 0.82B (β‰ˆ2.9 epochs)
FineWeb-Edu, English (sample-10BT, 5 of its files) HuggingFaceFW/fineweb-edu 4.42B (3.66M docs) 2.73B (β‰ˆ0.62 epochs)
  • Mixture. 88% FineWeb-2 / 2% Wikipedia / 10% English during the stable phase; 78% / 8% / 14% during the final learning-rate decay. The English share keeps some English ability and adds educational content.
  • Filtering. FineWeb-2 documents shorter than 200 characters (1.57M) or without Hangul (0.18M) were dropped.
  • Decontamination. Before any training, a held-out set was frozen: 2,000 FineWeb-2 test-split documents and 500 Wikipedia articles, the latter removed from training. Training documents were dropped if they shared any 13-word n-gram with a held-out document, or any 8-word n-gram with KoBEST (all splits), the KMMLU test set or our chat eval prompts. This removed 6.47M FineWeb-2 documents (10.6%), mostly template boilerplate shared with held-out pages.
  • Tokenizer. Byte-level BPE, 48,000 tokens, trained on a 2 GB sample (1.55 GB FineWeb-2, 0.25 GB Wikipedia, 0.2 GB English), NFC normalisation, single-digit splitting, 16 reserved special tokens (<|endoftext|>, chat-role tokens, …). 5.03 bytes per token on the held-out Korean text.

Training method

  • Objective. Next-token prediction on packed 2048-token sequences; each document is followed by <|endoftext|>.
  • Batch. 512 Γ— 2048 = 1.05M tokens per step (gradient accumulation), 24,000 steps, 25.17B tokens in total (β‰ˆ25 tokens per parameter).
  • Optimizer. Muon (Newton–Schulz-5 orthogonalised momentum, 0.95 Nesterov, with the RMS-matching update scale of Moonlight so one learning rate serves both optimizers) for all 2-D weight matrices; AdamW (Ξ² = 0.9/0.95) for embeddings, norms and scalar parameters. Weight decay 0.1, gradient clipping 1.0. In a short ablation (150M model) Muon reached 0.024 lower BPB than AdamW after 1,000 steps (0.5B tokens).
  • Learning-rate schedule. Warmup–stable–decay: linear warmup over 1,500 steps to 0.0015, constant until step 19,000, then a 1βˆ’sqrt decay to 0 over the last 5,000 steps (21% of training). Most of the late improvement comes in that decay: held-out FineWeb-2 BPB went from 0.786 at its start to 0.738 at the end (Wikipedia 0.837 β†’ 0.755).
  • Precision. bf16 autocast with fp32 master weights; the Gated DeltaNet triangular inverse and state recurrence run in fp32/TF32.
  • Hardware. NVIDIA RTX A6000 48 GB (PCIe), plain data parallel: 4 GPUs for the first 13k steps, then 2 GPUs (same global batch). About 680 GPU-hours.

validation BPB during pretraining

Evaluation

Held-out bits per byte (lower is better; tokenizer-independent, sliding window, stride 512) and KoBEST 0-shot accuracy, all models run through the same harness:

model BPB all / FineWeb-2 / Wikipedia COPA HellaSwag SentiNeg BoolQ WiC
this model (1.0B) 0.729 / 0.727 / 0.735 0.690 0.418 0.599 0.496 0.487
EleutherAI/polyglot-ko-1.3b 1.052 / 1.008 / 1.200 0.711 0.408 0.703 0.499 0.487
skt/kogpt2-base-v2 (125M) 1.092 / 1.073 / 1.157 0.584 0.318 0.615 0.502 0.490

KMMLU 0-shot: 0.255 (chance 0.25; zero-shot letter scoring is dominated by answer-letter bias at this scale, so it is reported for reference only).

Sample completions

One sample per prompt (seed 0, temperature 0.8, top-p 0.95, repetition penalty 1.1, up to 120 new tokens), unedited. The completions are fluent but contain factual errors (e.g. wrong population figures and an invented history of Hangul), so do not read them as facts.

λŒ€ν•œλ―Όκ΅­μ˜ μˆ˜λ„λŠ” λŒ€ν•œλ―Όκ΅­μ˜ μˆ˜λ„μ΄λ‹€.
λŒ€ν•œλ―Όκ΅­
λŒ€ν•œλ―Όκ΅­μ˜ μˆ˜λ„(首都) λ˜λŠ” μ•½μΉ­μœΌλ‘œ μˆ˜λ„λ₯Ό κ·œμ •ν•˜λŠ” 법령은 μ§€λ°©μžμΉ˜λ²• 제164쑰와 동법 μ‹œν–‰λ Ή 제32μ‘°λ‹€. 단, μˆ˜λ„μ˜ κ°œλ…μ— ν¬ν•¨ν•˜μ§€ μ•ŠλŠ” λ„μ‹œλ„ λ§Žλ‹€. 이듀 λ„μ‹œλ“€μ€ κ΄‘μ—­μžμΉ˜λ‹¨μ²΄λ₯Ό μ„€μΉ˜ν•˜κ±°λ‚˜ κ΄‘μ—­μ‹œκΈ‰ ν–‰μ •λ‹¨μœ„λ‘œμ„œ νŠΉλ³„μ‹œΒ·κ΄‘μ—­μ‹œΒ·νŠΉλ³„μžμΉ˜μ‹œ 및 도 등이 μ‘΄μž¬ν•œλ‹€.

μΈμ²œκ΄‘μ—­μ‹œλŠ” λŒ€ν•œλ―Όκ΅­ μ΅œλŒ€μ˜ λ„μ‹œμ΄λ‹€. 70만 λͺ… 정도가 κ±°μ£Όν•˜κ³  있으며 μˆ˜λ„κΆŒ 인ꡬ의 μ•½ μ ˆλ°˜μ„ μ°¨μ§€ν•˜κ³  μžˆλ‹€. λŒ€ν•œλ―Όκ΅­ 5λŒ€ κ΄‘μ—­μ‹œ μ€‘μ˜ ν•˜λ‚˜μ΄λ‹€.

μ„œμšΈνŠΉλ³„μ‹œμ—λŠ” 2천만 λͺ…이 κ±°μ£Όν•˜λ©° 주둜 μ„œμšΈ

μ‘°μ„  μ‹œλŒ€μ˜ κ³Όκ±° μ œλ„λŠ” 크게 문과와 무과(ζ­¦η§‘)둜 κ΅¬λΆ„λ˜μ—ˆλŠ”λ°, μ‘°μ„  μ „κΈ° κ΄€μ œμ˜ κ°€μž₯ 큰 νŠΉμ§•μ€ κ³Όκ±° μ œλ„κ°€ μ§€λ°©κ΄€μ˜ μΆœμ„Έ μˆ˜λ‹¨μœΌλ‘œ μ‚¬μš©λ˜μ—ˆλ‹€λŠ” 사싀이닀. μ΄λŠ” 지방관이 μ€‘μ•™μ˜ λΆ€μ„œλ₯Ό μž₯μ•…ν•¨μœΌλ‘œμ¨ μ€‘μ•™μ§‘κΆŒμ˜ 체제λ₯Ό κ΅¬μΆ•ν•˜κ³ μž ν–ˆλ˜ 것과 λ°€μ ‘ν•œ 관련이 μžˆλ‹€. κ·Έ κ²°κ³Ό, μ§€λ°©μ—λŠ” μ€‘μ•™μ—μ„œ μž„λͺ…λœ 관리가 파견됨으둜써 μ€‘μ•™κ³Όμ˜ 유기적 관계가 ν˜•μ„±λ˜μ—ˆλ‹€.
λ¬΄κ³ΌλŠ” μ „λž€μ΄ λΉˆλ²ˆν–ˆλ˜ ν›„κΈ°λΆ€ν„° 쇠퇴기에 걸쳐 μ‹œν–‰λ˜μ—ˆμœΌλ©°, κ΅­κ°€μ—μ„œ ν•„μš”λ‘œ ν•˜λŠ” μ „λ¬Έ 인λ ₯을 μ–‘μ„±ν•˜λŠ” 데 λͺ©μ μ΄ μžˆμ—ˆλ‹€. 무과의 μ£Όμš” κ³Όλͺ©μ€ μœ‘ν•΄κ³΅ 각 3개 κ³Όλͺ©μœΌλ‘œ κ΅¬μ„±λ˜μ—ˆκ³ , κΈ‰μ œν•˜λ©΄ 무과에 μ‘μ‹œ

광합성은 식물이 νƒœμ–‘ μ—λ„ˆμ§€λ₯Ό ν‘μˆ˜ν•˜μ—¬ νƒ„μˆ˜ν™”λ¬Όμ„ ν•©μ„±ν•˜λŠ” 과정이닀. κ΄‘ν•©μ„± κ³Όμ •μ—μ„œ 식물 μ„Έν¬λŠ” 엽둝체λ₯Ό 톡해 κ΄‘μ „μž μ—λ„ˆμ§€μΈ μ „μž 수용체(electron acceptor)λ₯Ό ν‘μˆ˜ν•œλ‹€.

μ „μžμˆ˜μš©μ²΄κ°€ ν¬ν•¨λœ λ¬Ό λΆ„μžλ₯Ό ν‘μˆ˜ν•œ 이후, μ—½λ‘μ²΄λŠ” μ—½λ‘μ²΄μ—μ„œ λ°©μΆœλ˜λŠ” κ΄‘μž μ—λ„ˆμ§€μ›μœΌλ‘œλΆ€ν„° μ—λ„ˆμ§€λ₯Ό μ–»λŠ”λ‹€. κ΄‘μžλ₯Ό ν‘μˆ˜ν•˜μ—¬ μ „μžκ°€ κ²°ν•©ν•˜κ²Œ 되면 이 결합을 μ „κΈ°μ μœΌλ‘œ 쀑성인 μ „μž μˆ˜μš©μ²΄μ™€ κ³΅μœ ν•¨μœΌλ‘œμ¨ μ „μžμ™€ κ²°ν•©λœ μ—λ„ˆμ§€κ°€ λ‚΄λΆ€ 합성을 ν†΅ν•œ 무기 인산 μƒμ„±μœΌλ‘œ μ „ν™˜λœλ‹€. μƒμ„±λœ 유기물이 λ°”λ‘œ νƒ„μ†Œμ΄λ‹€. λ”°λΌμ„œ κ΄‘ν•©μ„±μ˜ 핡심은 빛에 μ˜ν•΄ μ „

κΉ€μΉ˜μ°Œκ°œλ₯Ό λ§›μžˆκ²Œ 끓이렀면 λ¨Όμ € 재료λ₯Ό μ€€λΉ„ν•΄μ•Ό ν•œλ‹€.
μš°μ„  λŒ€νŒŒμ™€ μ–‘νŒŒλ₯Ό 썰고, 돼지고기 λͺ©μ‚΄μ€ ν•œ μž… 크기둜 자λ₯΄κ³  , λ„ν†°ν•œ μ•žλ‹€λ¦¬ μ‚΄κ³Ό 뢈고기감 μ†Œκ³ κΈ°λ₯Ό μ€€λΉ„ν•œλ‹€.
κΉ€μΉ˜λŠ” 무우 λ°°μΆ”μ˜ 3잎 정도, λ§›μˆ μ΄λ‚˜ μ†Œμ£Ό 약간을 μ„žμ–΄μ„œ 썰어쀀닀. 그리고 κ³ κΈ°λŠ” μ•žλ‹€λ¦¬μ‚΄μ„ μ€€λΉ„ν•˜κ³ , 기름을 κ±·μ–΄λ‚΄κ³  λ¨ΉκΈ° 쒋은 크기둜 잘라 λ†“λŠ”λ‹€. 냄비에 κ³ κΈ°λ₯Ό λ„£κ³ , 물을 λΆ“κ³  된μž₯ 반 ν°μˆ μ„ λ„£κ³  ν‘Ή μ‚Άμ•„ μ€€λ‹€. μ°Œκ°œκ°€ 끓기 μ‹œμž‘ν•˜λ©΄ λ‹€μ‹œλ§ˆ μœ‘μˆ˜μ™€ κ³ μΆ”μž₯으둜 간을 λ§žμΆ˜λ‹€.
그리고 돼지고기와 잘 읡은 κΉ€μΉ˜λ„ ν•¨κ»˜

였늘 μ„œμšΈμ€ μ•„μΉ¨λΆ€ν„° λΉ„κ°€ λ‚΄λ¦¬λ©΄μ„œ λŒ€κΈ°κ°€ λΆˆμ•ˆμ •ν•΄μ Έ 곳곳의 λ„λ‘œκ°€ κ³³κ³³μ—μ„œ ν†΅μ œλ˜κ³  μžˆμŠ΅λ‹ˆλ‹€.
였늘 μ˜€ν›„ 4μ‹œμ―€λΆ€ν„° 인천 κ°•ν™”, 경기도 김포 λ“± μˆ˜λ„κΆŒκ³Ό μ„œν•΄μ•ˆ 지역에 호우 νŠΉλ³΄κ°€ λ°œνš¨λμŠ΅λ‹ˆλ‹€.
특히 μ„œμšΈκ³Ό κ²½κΈ° 남뢀에 μ‹œκ°„λ‹Ή μ΅œλŒ€ 20mm μ•ˆνŒŽμœΌλ‘œ κ°•ν•˜κ²Œ λ‚΄λ¦¬λŠ” 곳이 μžˆκ² μŠ΅λ‹ˆλ‹€.
또 비ꡬ름 λ’€λ‘œ μ‚°λ°œμ μΈ λΉ—λ°©μšΈμ΄ λ–¨μ–΄μ§€λŠ” 곳도 μžˆκ² κ³ μš”.
κ³³κ³³μ—μ„œ 천λ‘₯ 번개λ₯Ό λ™λ°˜ν•œ κ°•ν•œ μ†Œλ‚˜κΈ°λ„ μŸμ•„μ§ˆ κ²ƒμœΌλ‘œ λ³΄μž…λ‹ˆλ‹€.
μ˜ˆμƒλ˜λŠ” κ°•μš°λŸ‰μ€ μ„œμšΈ λ‚¨μ„œ λ‚΄λ₯™μ—μ„œ λ§Žκ²ŒλŠ” 150mm 이상 μ˜ˆμƒλ©λ‹ˆλ‹€.
λΉ„μ˜ 양은 쀑뢀와 남뢀 μ§€λ°© λͺ¨λ‘ 30~

인곡지λŠ₯ 기술이 λ°œμ „ν•˜λ©΄μ„œ μΌμžλ¦¬μ— λ―ΈμΉ˜λŠ” 영ν–₯은 λ”μš± 컀지고 μžˆμŠ΅λ‹ˆλ‹€. 인곡지λŠ₯이 인간을 λŒ€μ²΄ν•  κ²ƒμ΄λΌλŠ” 전망이 λ‚˜μ˜€λŠ” κ°€μš΄λ°, μ΄λŸ¬ν•œ 변화에 λ§žμ„œκΈ° μœ„ν•œ λ…Έλ ₯이 이루어지고 μžˆμŠ΅λ‹ˆλ‹€. 특히, λ―Έλž˜μ—λŠ” 인곡지λŠ₯ 기술이 더 λ°œμ „ν•  κ²ƒμœΌλ‘œ μ˜ˆμƒλ˜λ©΄μ„œ 이에 λŒ€ν•œ λŒ€λΉ„κ°€ λ”μš± μ€‘μš”ν•΄μ§€κ³  μžˆλŠ”λ°μš”. κ·Έλ ‡λ‹€λ©΄ λ―Έλž˜μ— λŒ€λΉ„ν•˜λŠ” κ΅μœ‘μ€ μ–΄λ–€ λ°©ν–₯으둜 μ§„ν–‰λ κΉŒμš”? 였늘 IBKκΈ°μ—…μ€ν–‰μ—μ„œλŠ” 미래 μ‚¬νšŒμ˜ μΈμž¬μƒμ„ κ³ λ €ν•œ ꡐ윑 λ°©ν–₯성을 μ‚΄νŽ΄λ³΄κ² μŠ΅λ‹ˆλ‹€. 4μ°¨ μ‚°μ—…ν˜λͺ… μ‹œλŒ€λ₯Ό μ‚΄μ•„κ°€λŠ” μš°λ¦¬μ—κ²Œ ν•„μš”ν•œ μ—­λŸ‰μ€ λ¬΄μ—‡μΌκΉŒμš”? μ˜€λŠ˜μ€ 미래λ₯Ό μ€€λΉ„ν•˜λŠ” κ΅μœ‘μ—μ„œ μ€‘μš”ν•˜κ²Œ μƒκ°ν•˜λŠ” μ—­λŸ‰μ„ μ‚΄νŽ΄λ³΄λ €κ³  ν•©λ‹ˆλ‹€. β–  4μ°¨μ‚°μ—…ν˜λͺ…..
금육의 μƒˆλ‘œμš΄ μ˜μ—­μΈ ν•€ν…Œν¬ μŠ€νƒ€νŠΈμ—…! ν•€ν…Œν¬ 산업은 금육과 IT기술의 μœ΅ν•©μœΌλ‘œ νƒ„μƒν•œ μƒˆλ‘œμš΄ μ‹œμž₯으둜, κΈ°μ‘΄ 은행 및

The history of the Korean alphabet, Hangul, begins with the vowel-numbering system in 1440 CE (Korean: Kyeonggisan Jujipunsil). The first consonant-vowel-picture script was the alphabetic writing system introduced by Samgak (Samgyecheok) and is based on the nodal point between throat and tongue. The first important stage in the development of Korean alphabetic writing system began when the Joseon Dynasty (1392–1897) changed from

Usage

Requires a CUDA GPU, PyTorch, Triton, tokenizers and safetensors (tested with PyTorch 2.8 + Triton 3.4). The model is not a transformers architecture; load it with the included code:

import sys
from huggingface_hub import snapshot_download
path = snapshot_download("shopkeeper/kor-finewiki-1b-base")
sys.path.insert(0, path)
from kollm.load import load_model
from kollm.generate import generate, special_ids

model, tok, _ = load_model(path)                       # bf16 on cuda
eot = special_ids(tok)["endoftext"]
ids = [eot] + tok.encode("λŒ€ν•œλ―Όκ΅­μ˜ μˆ˜λ„λŠ”").ids        # documents start after <|endoftext|>
out = generate(model, ids, max_new_tokens=100, temperature=0.8, top_p=0.95, stop_ids={eot})
print(tok.decode(out))

Or interactively: python complete.py inside the downloaded folder.

License and data attribution

Model weights and code: apache-2.0. Training data: FineWeb-2 and FineWeb-Edu (ODC-By 1.0, subject to Common Crawl's terms of use) and Wikipedia (CC BY-SA 4.0).

Downloads last month
430
Safetensors
Model size
1B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for shopkeeper/kor-finewiki-1b-base

Finetunes
1 model

Datasets used to train shopkeeper/kor-finewiki-1b-base