max-0.1b v1
A small Korean-first base language model (109.5M parameters) for industrial small functions such as document classification, request routing, field extraction and form filling. v1 continues v0 from its pre-decay checkpoint for 7.6B more tokens (10.0B in total). This is a base model (not instruction-tuned): a starting point for task fine-tuning and research, not for direct end-user use.
한국어 중심의 소형 기초 언어모델(1.095억 매개변수)입니다. 문서 분류, 요청 분기, 항목 추출, 양식 작성 같은 산업용 작은 기능의 출발점으로 쓰는 기초 모델입니다. 지시 따르기 미세조정은 하지 않았습니다. v0의 감쇠 전 저장분에서 이어서 학습했고, 누적 학습량은 100억 토큰입니다.
Files
| File | Use |
|---|---|
model.safetensors, config.json, tokenizer.json |
Transformers (Qwen3ForCausalLM-compatible layout) |
gguf/max01b-v1-f16.gguf (221 MB) |
llama.cpp, full precision (greedy output identical to Transformers on 5/5 test prompts) |
gguf/max01b-v1-q4_k_m.gguf (73 MB) |
llama.cpp, 4-bit, runs on ordinary CPUs (perplexity about 8% higher than f16) |
Context length: 1,024 tokens. Vocabulary: 32,000 (Korean-optimized, digits split; same tokenizer as v0). Input text should be Unicode NFC-normalized (NFD Korean inflates the token count about 10x).
llama-completion -m max01b-v1-q4_k_m.gguf -p "오늘 회의에서 논의된 안건은" -n 64 -no-cnv
from transformers import AutoModelForCausalLM, PreTrainedTokenizerFast
tok = PreTrainedTokenizerFast.from_pretrained("quantumaikr/max-0.1b-v1")
model = AutoModelForCausalLM.from_pretrained("quantumaikr/max-0.1b-v1")
Training
- Architecture as v0: 12 layers, width 768, SwiGLU 2048, RoPE, QK-norm, tied embeddings (Qwen3-compatible names).
- Continued from v0
stable_pre_decay(2.40B tokens) for 7.60B tokens on 2 GPUs (DDP); last 20% linear learning-rate decay. - Optimizer: Muon (hidden matrices) + AdamW (embeddings, scalars). Batch 524,288 tokens, context 1,024.
- Data (all ODC-By 1.0): FineWeb-2 Korean, FinePDFs Korean, FineWeb-Edu (incl. score ≥ 4 subset), FineMath-4plus. Mix: Korean web 40% / Korean documents 20% / English educational 33% / math 7% (decay phase: 33 / 25 / 32 / 10). Cleaning: NFC, script-ratio and length filters, exact deduplication, regex replacement of e-mail, phone, account and resident-registration-number patterns. No code data, no B/C-grade corpora, no model-generated text.
Evaluation
Validation loss (held-out, same files as v0; lower is better):
| set | v0 (2.8B tokens) | v1 (10.0B) |
|---|---|---|
| Korean web | 3.199 | 3.055 (0.903 bpb) |
| English | 2.961 | 2.851 (1.020 bpb) |
| Korean documents (PDF) | 2.341 | 2.208 (0.791 bpb) |
| Math | 2.185 | 2.020 (0.936 bpb) |
Fine-tuned on public, human-written Korean tasks (train split, 2 epochs; LR chosen on dev; test mean ± std over 3 seeds):
| task | v0 | v1 | KLUE-RoBERTa-base (ref.) |
|---|---|---|---|
| KLUE-YNAT (macro F1) | 84.4 ± 0.3 | 85.1 ± 0.0 | 86.4 |
| KLUE-NLI (acc) | 70.4 ± 0.6 | 73.8 ± 1.0 | 85.1 |
| KLUE-NER (entity F1) | 81.1 ± 0.3 | 82.5 ± 0.4 | 87.4 |
| NSMC (acc) | 86.4 | 87.1 ± 0.1 | 88.4 |
| MASSIVE ko intent (acc) | 86.4 ± 0.2 | 86.1 ± 0.4 | 86.5 |
| MASSIVE ko slot (F1) | 74.6 ± 0.1 | 74.9 ± 0.3 | 74.6 |
| 3i4K (acc) | 89.6 ± 0.0 | 89.5 ± 0.3 | 90.0 |
The decoder is generative (answers are generated text scored by the same scorer for every model); the encoder reference uses classification/tagging heads on the same splits.
Limitations
- Small base model: limited world knowledge and reasoning; it can produce fluent but wrong text. Fine-tune before use.
- Personal-information patterns were replaced with regular expressions; names and other free-form personal data may remain in the training text. Do not use the model to extract personal data.
- Korean and English only; context 1,024 tokens.
License
Apache-2.0 for the model weights. Training data: FineWeb-2, FinePDFs, FineWeb-Edu and FineMath (Hugging Face), ODC-By 1.0; please keep the attribution when redistributing derived datasets.
Training checkpoints
PyTorch training checkpoints (moved here from GitHub releases):
| file | what |
|---|---|
checkpoints/max01b-v1-stable_pre_decay_v1.pt |
stable phase at full LR, includes optimizer state (resume pretraining) |
checkpoints/max01b-v1-final.pt |
final checkpoint after cooldown |
Verify with checkpoints/SHA256SUMS.