You can deploy this model and start earning money today!

Ethualo/
Qwen3-1.7b-korean-compressor-GGUF

Name: Qwen3-1.7b-korean-compressor-GGUF
Availability: InStock

Views

No views yet

Qwen3-1.7B-korean-compressor-GGUF

Qwen3-1.7B fine-tuned for ko-prompt-compressor. A single-purpose model that compresses Korean prompts to 30-50% of their original length while preserving meaning.

Training

Base: unsloth/Qwen3-1.7B
Method: QLoRA (rank 16, alpha 32, target: q/k/v/o/gate/up/down_proj) via Unsloth
Data: 4,000-row SFT distillation dataset, filtered from compression examples generated by three teacher models: Haiku 4.5, gemma-4-E2B-it-qat and Hy-MT2-7B
Context length: 1024
Optimizer: AdamW 8-bit, linear LR schedule, lr 2e-4, batch size 8 x grad accum 4

Usage

python

from llama_cpp import Llama

llm = Llama(
    model_path="Qwen3-1.7B-korean-compressor-GGUF.gguf",
    n_ctx=1024,
    n_gpu_layers=-1,  # offload all layers to GPU; use 0 for CPU-only
)

SYSTEM_PROMPT = "Compress the Korean prompt. Output ONLY the compressed question. ..."

out = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": "Original Korean prompt"},
    ],
    max_tokens=256,
    temperature=0.1,
)
print(out["choices"][0]["message"]["content"])

Auto-wired through the ko-prompt-compressor package:

bash

pip install "ko-prompt-compressor[llm]"
ko-compress "original prompt" --model qwen3-ko-compressor

Verification

Compression output is meant to be checked separately by ko-prompt-compressor's verifier (critical_guard hard-veto rules + e5-large-korean embedding similarity). Without the verifier, this model's output is not guaranteed to preserve meaning.

License

Apache 2.0, inherited from the base model (Qwen3-1.7B).

Ethualo/Qwen3-1.7b-korean-compressor-GGUF