CLinker

The CLinker models are distilled language models specifically designed for command-line graph construction, developed by CyCraft AI Lab. CLinker was instroduced in SINCON 2025, with talk titled "CLINKER — An Efficient Distilled LLM Command Line Graph Constructor".

Usage

Launch openai-compatible server (e.g., vllm)

python3 -m vllm.entrypoints.openai.api_server \
    --host 0.0.0.0 \
    --port 3000 \
    --served-model-name $model_name \
    --max-model-len $length \
    --api-key $api_key \
    --model $model_path

DSPy inference

import dspy

# Set dspy module default LM
lm = dspy.LM(
    model=f'openai/{$model_name}',
    api_key=f'{$api_key}',
    api_base='http://localhost:3000/v1',
    model_type='chat',
    temperature=0.7,
    max_tokens=4000,
    cache=False,
    num_retries=0
)
dspy.configure(lm=lm)

from command_parser import CmdlineParser, CoTCmdlineParser
from command_extractor import CmdlineExtractor, CoTCmdlineExtractor

cmdline = 'echo hello world'

# Reasoning model `CLinker-DeepSeek-1.5B` use non-chain-of-thought prompt
parser = CmdlineParser()
extractor = CmdlineExtractor()
# Non-reasoning models are equipped with chain-of-thoughts prompt
parser = CoTCmdlineParser()
extractor = CoTCmdlineExtractor()

# Run inference
parser_response = parser(cmdline).toDict()
extractor_response = extractor(cmdline).toDict()

# Transform Response: pydantic.BaseModel object into dict
parser_response['response'] = parser_response['response'].model_dump(mode='json')

print(parser_response)
print(extractor_response)

Graph construction

from command_graph_builder import build_cmdline_graph

graph: nx.DiGraph = build_cmdline_graph(cmdline, parser_response, extractor_response)