Context Segmentation for SLM Agents

This is the implementation of the solution proposed in the scientific paper: Evaluating Context Segmentation in Locally Deployable SLMs for Cybersecurity CTF Tasks

Setup

Ensure you have Python 3.12.3+ installed along with Docker engine

Python environment

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r requirements.txt

llama.cpp

Start llama.cpp with Vulkan

docker compose --profile vulkan up -d

or with Cuda

docker compose --profile cuda up -d

Build Models and Containers

Download your desired Large Language Model (LLM) in GGUF format from HuggingFace, then build the intercode-ctf containers.

docker exec -it llamacpp ./llama-cli -hf unsloth/gemma-4-E2B-it-GGUF:Q4_K_M  # or any other GGUF model on HF
docker build -t intercode-ctf -f intercode/docker/ctf.Dockerfile ./intercode
docker restart llamacpp

Run benchmark

Run N benchmarks of a strategy (plain or explorer)

This accumulates results into a $MODEL_NAME.$STRATEGY.bench file

OPENAI_API_BASE=http://127.0.0.1:8080/ \
OPENAI_API_KEY=1234 \ #Mock key (we run locally)
python3 intercode_test.py \
  --model unsloth/gemma-4-E2B-it-GGUF:Q4_K_M \ #HF model
  --strategy plain \ #plain or explorer
  --count 1 #N

Check benchmark results

jq '.' gemma-4-E2b-it-GGUF:Q4_K_M.plain.bench

Remarks

Code in the directory intercode was taken from princeton-nlp/intercode and adjusted to make it work for our setup.