HybriKo-117M-LinuxFC-SFT-v2

Korean Hybrid LLM fine-tuned for Linux Command Function Calling.

πŸš€ 20초 λ§Œμ— Colabμ—μ„œ μ‹€ν–‰ κ°€λŠ₯! GPU T4μ—μ„œ λ°”λ‘œ ν…ŒμŠ€νŠΈν•΄λ³΄μ„Έμš”.

Architecture

Griffin-style Hybrid Architecture - RNNκ³Ό Attention을 2:1 λΉ„μœ¨λ‘œ κ²°ν•©ν•˜μ—¬ νš¨μœ¨μ„±κ³Ό μ„±λŠ₯을 λ™μ‹œμ— ν™•λ³΄ν•©λ‹ˆλ‹€.

Component Description
Griffin Block RG-LRU (Real-Gated Linear Recurrent Unit) + GeGLU FFN
Attention Block GQA (Grouped Query Attention) + RoPE + GeGLU FFN
Pattern Griffin β†’ Griffin β†’ Attention (2:1 반볡)
Total Layers 12 (Griffin 8 + Attention 4)

Model Details

Spec Value
Parameters 117.8M
d_model 768
n_layers 12
n_heads 12
n_kv_heads 3
max_seq_len 6,144
vocab_size 32,000

Training Results

ν΄λ‘œλ“œμ™€ μ œλ―Έλ‚˜μ΄λ‘œ 5,250개 ν•™μŠ΅ μƒ˜ν”Œμ„ μƒμ„±ν•˜μ—¬ λŒ€ν­ ν–₯μƒλœ μ„±λŠ₯을 λ‹¬μ„±ν–ˆμŠ΅λ‹ˆλ‹€.

Metric Before (250 samples) After (5K samples)
Train Samples 250 4,725
Eval Loss 0.039 0.0039
Action Name Accuracy 4% (2/50) 100% (100/100)

κ²°λ‘ : 데이터 증가 (250 β†’ 5,000)둜 4% β†’ 100% Action Name 정확도 달성!

⚠️ Known Limitations

  • Action Name: 100% μ •ν™• (μ˜¬λ°”λ₯Έ λͺ…λ Ήμ–΄ 선택)
  • Parameters: 일뢀 hallucination λ°œμƒ κ°€λŠ₯
    • 파일λͺ…/κ²½λ‘œκ°€ λ‹€λ₯΄κ²Œ 생성될 수 있음 (예: test.txt β†’ test.py)
    • Thoughtκ°€ μ˜λ„μ™€ λ‹€λ₯Ό 수 있음

117M νŒŒλΌλ―Έν„°μ˜ μ†Œν˜• λͺ¨λΈ ν•œκ³„μž…λ‹ˆλ‹€. 더 μ •ν™•ν•œ νŒŒλΌλ―Έν„° 생성을 μœ„ν•΄μ„œλŠ” 더 큰 λͺ¨λΈκ³Ό 더 λ§Žμ€ 데이터가 ν•„μš”ν•©λ‹ˆλ‹€.

Supported Commands (21)

ls, cd, mkdir, rm, cp, mv, find, cat, grep, head, 
tail, wc, ps, df, du, top, ping, curl, chmod, tar, Finish

Quick Start (Colab)

!pip install -q huggingface_hub sentencepiece
from huggingface_hub import hf_hub_download
hf_hub_download("Yaongi/HybriKo-117M-LinuxFC-SFT-v2", "demo_colab.py", local_dir=".")

!python demo_colab.py --query "λ””μŠ€ν¬ μ‚¬μš©λŸ‰μ„ ν™•μΈν•΄μ€˜"
!python demo_colab.py --query "test.txt νŒŒμΌμ„ μ‚­μ œν•΄μ€˜"
!python demo_colab.py --query "backup.tar.gz μ••μΆ• ν’€μ–΄μ€˜"
!python demo_colab.py --query "ν˜„μž¬ ν΄λ”μ˜ 파일 λͺ©λ‘μ„ λ³΄μ—¬μ€˜"

Example Outputs (Actual Results)

Input: λ””μŠ€ν¬ μ‚¬μš©λŸ‰μ„ ν™•μΈν•΄μ€˜
----------------------------------------
Thought: λ””μŠ€ν¬ μš©λŸ‰μ„ ν™•μΈν•©λ‹ˆλ‹€.
Action:  df_command
Input:   {"options": "-h --total"}
----------------------------------------

Input: test.txt νŒŒμΌμ„ μ‚­μ œν•΄μ€˜
----------------------------------------
Thought: rm λͺ…λ Ήμ–΄λ‘œ μ‚­μ œν•©λ‹ˆλ‹€.
Action:  rm_command
Input:   {"path": "test.py", "recursive": false}  ⚠️ 파일λͺ… hallucination
----------------------------------------

Input: backup.tar.gz μ••μΆ• ν’€μ–΄μ€˜
----------------------------------------
Thought: νŒŒμΌλ“€μ„ λ¬Άμ–΄μ„œ μ••μΆ•ν•©λ‹ˆλ‹€.  ⚠️ Thought 였λ₯˜ (μ‹€μ œλ‘œλŠ” μ••μΆ• ν•΄μ œ)
Action:  tar_command
Input:   {"options": "-xzf", "archive": "data.tar.gz"}  ⚠️ 파일λͺ… hallucination
----------------------------------------

Input: ν˜„μž¬ ν΄λ”μ˜ 파일 λͺ©λ‘μ„ λ³΄μ—¬μ€˜
----------------------------------------
Thought: ls λͺ…λ Ήμ–΄λ‘œ 파일 λͺ©λ‘μ„ λ΄…λ‹ˆλ‹€.
Action:  ls_command
Input:   {"path": "/etc", "options": "-lS"}  ⚠️ 경둜 hallucination
----------------------------------------

Note: Action Name (df_command, rm_command, tar_command, ls_command)은 λͺ¨λ‘ μ •ν™•ν•©λ‹ˆλ‹€.

Training Details

Config Value
Hardware A100 x 8 (DDP)
Batch Size 32 (1 Γ— 8 GPUs Γ— 4 grad accum)
Learning Rate 5e-5
Warmup Steps 100
Epochs 15
Base Model HybriKo-117M (exp7_phase1)

Files

File Description
pytorch_model.pt Model weights
HybriKo_tok.model SentencePiece tokenizer
demo_colab.py Colab demo script (auto-downloads all files)
configuration_hybridko.py Model config
modeling_hybridko.py Model implementation

License

Apache 2.0

Citation

@misc{hybridko-linuxfc-2026,
  title={HybriKo-117M-LinuxFC-SFT-v2: Korean Hybrid LLM for Linux Function Calling},
  author={Yaongi},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/Yaongi/HybriKo-117M-LinuxFC-SFT-v2}
}
Downloads last month
7
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support