How to use from
llama.cpp
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf B1NARIES/ShellAssist-Qwen3-0.6B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf B1NARIES/ShellAssist-Qwen3-0.6B:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf B1NARIES/ShellAssist-Qwen3-0.6B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf B1NARIES/ShellAssist-Qwen3-0.6B:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf B1NARIES/ShellAssist-Qwen3-0.6B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf B1NARIES/ShellAssist-Qwen3-0.6B:Q4_K_M
Use Docker
docker model run hf.co/B1NARIES/ShellAssist-Qwen3-0.6B:Q4_K_M
Quick Links

ShellAssist Qwen3-0.6B

RU/EN shell-command assistant fine-tuned on 30,000 instructionโ†’command records with QLoRA 4-bit SFT (3 epochs) followed by DPO (216 steps).

Model

  • Base: Qwen/Qwen3-0.6B
  • Method: QLoRA nf4 SFT + DPO, merged fp16 weights and GGUF (f16 / q8_0 / q4_k_m)
  • Training data: B1NARIES/shellassist-sft-dpo (private, 30k records, exact platform/type/language quotas)
  • Eval-300 (authored tasks): 16/300 exact (5.3%); internal test 88/120 (73.3%)
  • Think blocks (TASK/ENV/PLAN/RISK/CHECK) are present in ~75% of training records.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")
model = PeftModel.from_pretrained(base, "B1NARIES/ShellAssist-Qwen3-0.6B")
tok = AutoTokenizer.from_pretrained("B1NARIES/ShellAssist-Qwen3-0.6B")
# enable_thinking=True, then extract the fenced command from the answer

Limitations

The model was trained on template-style instructions; generalisation to unseen phrasings is limited (see eval-300 above). Platform is not always named in the user turn. Not suitable for destructive operations without human review.

Author

B1NARIES (Binar gg)

Downloads last month
151
GGUF
Model size
0.6B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for B1NARIES/ShellAssist-Qwen3-0.6B

Finetuned
Qwen/Qwen3-0.6B
Adapter
(769)
this model