qwen-ctfer

qwen-ctfer is a QLoRA adapter for Qwen/Qwen2.5-Coder-7B-Instruct tuned on transformed CTF writeups.

What This Is

• Base model: Qwen/Qwen2.5-Coder-7B-Instruct
• Fine-tuning method: QLoRA
• Scope: concise CTF-style solve paths, exploit notes, debugging steps, and operator-focused walkthroughs
• Artifact: adapter weights, not merged full weights

Data

Training used authorized writeup-derived records built from:

• CTFtime writeup listings
• linked external writeups
• GitHub writeups fetched via raw content when possible
• non-GitHub pages transformed through r.jina.ai

Before training, the pipeline:

• cached raw fetches locally
• preferred original writeups over duplicate CTFtime mirrors when available
• masked obvious flag patterns
• dropped some low-signal profile and listing pages

The resulting dataset for this run contained:

• train: 460 samples
• eval: 50 samples

Training

• Rank: r=16
• Alpha: 32
• Dropout: 0.05
• Max sequence length: 2048
• Epochs: 2
• Batch size: 1
• Gradient accumulation: 8
• Learning rate: 2e-4

Evaluation

This run completed local prompt eval against the repository's current data/eval/example_cases.jsonl.

• base model: 0/2
• tuned adapter: 0/2

This is an early publishable adapter, not a final benchmarked release.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-Coder-7B-Instruct"
adapter_id = "edamamex/qwen-ctfer"

tokenizer = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

Limitations

• This adapter can still hallucinate exploitability or tool output.
• Some transformed writeups still contain formatting noise from source platforms.
• All outputs should be validated against the actual target and rules of engagement.