HybriKo-117M-LinuxFC-SFT-v2
Korean Hybrid LLM fine-tuned for Linux Command Function Calling.
π 20μ΄ λ§μ Colabμμ μ€ν κ°λ₯! GPU T4μμ λ°λ‘ ν μ€νΈν΄λ³΄μΈμ.
Architecture
Griffin-style Hybrid Architecture - RNNκ³Ό Attentionμ 2:1 λΉμ¨λ‘ κ²°ν©νμ¬ ν¨μ¨μ±κ³Ό μ±λ₯μ λμμ ν보ν©λλ€.
| 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 λ°λ³΅) |
| 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‘λμ μ λ―Έλμ΄λ‘ 5,250κ° νμ΅ μνμ μμ±νμ¬ λν ν₯μλ μ±λ₯μ λ¬μ±νμ΅λλ€.
| 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‘ : λ°μ΄ν° μ¦κ° (250 β 5,000)λ‘ 4% β 100% Action Name μ νλ λ¬μ±!
β οΈ Known Limitations
- Action Name: 100% μ ν (μ¬λ°λ₯Έ λͺ λ Ήμ΄ μ ν)
- Parameters: μΌλΆ hallucination λ°μ κ°λ₯
- νμΌλͺ
/κ²½λ‘κ° λ€λ₯΄κ² μμ±λ μ μμ (μ:
test.txtβtest.py) - Thoughtκ° μλμ λ€λ₯Ό μ μμ
- νμΌλͺ
/κ²½λ‘κ° λ€λ₯΄κ² μμ±λ μ μμ (μ:
117M νλΌλ―Έν°μ μν λͺ¨λΈ νκ³μ λλ€. λ μ νν νλΌλ―Έν° μμ±μ μν΄μλ λ ν° λͺ¨λΈκ³Ό λ λ§μ λ°μ΄ν°κ° νμν©λλ€.
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 "λμ€ν¬ μ¬μ©λμ νμΈν΄μ€"
!python demo_colab.py --query "test.txt νμΌμ μμ ν΄μ€"
!python demo_colab.py --query "backup.tar.gz μμΆ νμ΄μ€"
!python demo_colab.py --query "νμ¬ ν΄λμ νμΌ λͺ©λ‘μ 보μ¬μ€"
Example Outputs (Actual Results)
Input: λμ€ν¬ μ¬μ©λμ νμΈν΄μ€
----------------------------------------
Thought: λμ€ν¬ μ©λμ νμΈν©λλ€.
Action: df_command
Input: {"options": "-h --total"}
----------------------------------------
Input: test.txt νμΌμ μμ ν΄μ€
----------------------------------------
Thought: rm λͺ
λ Ήμ΄λ‘ μμ ν©λλ€.
Action: rm_command
Input: {"path": "test.py", "recursive": false} β οΈ νμΌλͺ
hallucination
----------------------------------------
Input: backup.tar.gz μμΆ νμ΄μ€
----------------------------------------
Thought: νμΌλ€μ λ¬Άμ΄μ μμΆν©λλ€. β οΈ Thought μ€λ₯ (μ€μ λ‘λ μμΆ ν΄μ )
Action: tar_command
Input: {"options": "-xzf", "archive": "data.tar.gz"} β οΈ νμΌλͺ
hallucination
----------------------------------------
Input: νμ¬ ν΄λμ νμΌ λͺ©λ‘μ 보μ¬μ€
----------------------------------------
Thought: ls λͺ
λ Ήμ΄λ‘ νμΌ λͺ©λ‘μ λ΄
λλ€.
Action: ls_command
Input: {"path": "/etc", "options": "-lS"} β οΈ κ²½λ‘ hallucination
----------------------------------------
Note: Action Name (df_command, rm_command, tar_command, ls_command)μ λͺ¨λ μ νν©λλ€.
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