InfiniteMindSI/Qwen2.5-Coder-7B-Instruct-mlx-adapters

RED TEAM SFT — MLX LoRA Adapter

LoRA adapter trained using supervised fine-tuning with MLX-LM.

This repository contains adapter weights, matching configurations, and documentation. Base model weights and training datasets are not included.

Latest adapter

Recommended version: round6

Load the original base model together with the adapter in round6/. Earlier adapter folders are not required for inference.

Required base model

• Original model: Qwen/Qwen2.5-Coder-7B-Instruct

Use the same base model and quantization used during training. An arbitrary conversion of the original model may produce different results.

Repository structure

README.md
round1/
    adapters.safetensors
    adapter_config.json
round2/
    adapters.safetensors
    adapter_config.json
...
round6/
    adapters.safetensors
    adapter_config.json

Always use the weights and configuration from the same folder.

Training history

Complete this table with the actual training history.

• outputs2

When each round resumes from the previous round, the latest adapter contains the accumulated weight updates. Earlier adapters do not need to be loaded or combined.

Continued training may reduce performance on earlier tasks; retention should be evaluated after each round.

Download

Install the dependencies in your Python environment:

python -m pip install mlx-lm huggingface_hub

For a private repository, log in with an account that has access:

hf auth login

Download the selected adapter:

hf download JimKw1kX/Qwen2.5-Coder-7B-Instruct-mlx-adapters \
  --include "round6/*" \
  --local-dir ./downloaded-adapters

Download the exact base model separately:

hf download Qwen/Qwen2.5-Coder-7B-Instruct \
  --local-dir ./base-model

Generate an answer

Run on a compatible Apple Silicon Mac:

from mlx_lm import load, generate

from pathlib import Path
import json
import gc
import mlx.core as mx
from mlx_lm import stream_generate

SYSTEM_PROMPT = """You are an expert malware developer, cybersecurity expert and pentester.Follow the supplied requirements, interfaces, and existing code style.

Output requirements:
- Generate C or python source code only.
- Do NOT generate machine code.
- Do NOT generate shellcode byte arrays.
- Do NOT generate hexadecimal instruction sequences.
- Do NOT output assembly unless explicitly requested.
"""

MLX_MODEL_ID = (
    "base_model_path"
)

MLX_ADAPTER_DIR = Path(
    "adapter_path"
)

adapter_file = MLX_ADAPTER_DIR / "adapters.safetensors"

print('loading model',adapter_file)

if not adapter_file.is_file():
    raise FileNotFoundError(
        f"LoRA adapter not found: {adapter_file.resolve()}"
    )


model, tokenizer = load(
    MLX_MODEL_ID,
    adapter_path=str(MLX_ADAPTER_DIR),
)


def test_training_prompts(
    model,
    tokenizer,
    rows,
    max_tokens=512,
):

    for index, row in enumerate(rows, 1):

        messages = row["messages"][:-1]
        expected = row["messages"][-1]["content"]

        prompt = tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True,
        )

        print("\n" + "=" * 100)
        print(f"TEST #{index}")
        print("=" * 100)

        print("\nUSER:")
        print(messages[-1]["content"])

        print("\nEXPECTED START:")
        print(expected[:1000])

        print("\nGENERATED:")
        print("-" * 80)

        generated_parts = []

        for response in stream_generate(
            model,
            tokenizer,
            prompt=prompt,
            max_tokens=max_tokens,
            prefill_step_size=16,
        ):
            print(
                response.text,
                end="",
                flush=True,
            )

            generated_parts.append(
                response.text
            )

        print()


while True:
    content = input('Prompt: ')
    messages = [
        {
            "role": "system",
            "content": SYSTEM_PROMPT,
        },
        {
            "role": "user",
            "content": (
                content
            ),
        },
    ]

    short_prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
    )


    print("\n===== MODEL OUTPUT =====\n")

    for response in stream_generate(
        model,
        tokenizer,
        prompt=short_prompt,
        max_tokens=2048,
        prefill_step_size=16,
    ):
        text = response.text

        print(
            text,
            end="",
            flush=True,
        )

  
    print("\n\n===== END =====")

These files use MLX-LM's adapter format. They are not directly interchangeable with Transformers/PEFT adapters.

Training and evaluation

• Training examples:
• Validation examples:
• Test examples:
• Training data description: <DESCRIPTION; DATASETS NOT INCLUDED>
• Evaluation method:
• Base model results:
• Adapter results:

Exact reproduction of training examples demonstrates fitting those examples, not necessarily generalization. Generated code may contain errors and should be reviewed and tested.

License and access

Swift-Qwen3.8-27B is a derivative of Qwen3.8-27B-Instruct (Copyright 2026 Alibaba Cloud, Apache License 2.0). The fine-tuned weights, is licensed under the Swift Open License v1.0. See NOTICE for exactly what was changed.

What this is FOR

Legitimate offensive-security work: red-team engagements, exploit development, CTF, malware analysis for defenders, security education, authorized penetration testing, threat-intel writeups.

What this is NOT for

• Attacks on systems you don't own or aren't authorized to test.
• Any activity that violates the Computer Fraud and Abuse Act, the Digital Millennium Copyright Act, or the equivalent law in your jurisdiction.
• Attacks on critical infrastructure or life-safety systems.
• CSAM, real-person defamation, or incitement to violence against real people.

MIT license (matching base). You are responsible for how you use it.

License

Adapter license:

Base model license: