TobiasLogic/qwen-2.5-coder-Uncensored

Qwen2.5-Coder-1.5B — Uncensored Coding Fine-tune

A LoRA fine-tune of Qwen/Qwen2.5-Coder-1.5B (the base model, not Instruct), trained to behave as a direct, compliant coding assistant for software engineering, application security, and pentest-tooling tasks — without the refusal and disclaimer behavior typical of instruction-tuned chat models.

Trained by Htfi / TobiasLogic on a single RTX 3050 Laptop (4GB) using Unsloth.

What this is (plainly)

This model is uncensored for coding tasks. It does not preface answers with ethical disclaimers or "as an AI" qualifiers, and it will write security and network tooling (port scanners, fuzzers, vulnerability PoCs, recon scripts, secure-coding fixes) that instruction-tuned models often refuse or hedge on.

The "uncensored" behavior comes entirely from the training data — a synthetic dataset whose outputs contain no refusals — applied on top of a base model that has no built-in refusal layer. It is not a jailbreak; it is a model that simply never learned to refuse.

Intended use

Built for authorized work: writing and reviewing code, security auditing of systems you own or are permitted to test, CTFs, lab environments, and learning. The pentest-tooling capability is intended for legitimate, authorized security testing — the same category of tools taught in any security course and shipped in distributions like Kali.

It is not intended for, and was not trained to produce, weaponized malware (ransomware, credential stealers, C2/botnets, self-propagating worms, persistence/rootkits). Use it lawfully and only against targets you have permission to test.

Capabilities and limits

• Strong at: self-contained scripts and functions across Python, JavaScript, Go, and others; secure-coding fixes; standard recon/scanning/web-testing tooling.
• Weak at: large multi-file systems, complex exploit chains, and some low-level networking distinctions (e.g. it may conflate a TCP connect-scan with a true raw-socket SYN scan). This is the capability ceiling of a 1.5B model, not a training defect.
• Always review and test generated code before running it. Generated security tooling can be destructive or noisy on a network — run it only where you're authorized to.

Training details

• Base: Qwen/Qwen2.5-Coder-1.5B (base)
• Method: QLoRA (4-bit), rank 16, all linear layers, completion-only loss
• Data: ~1,200 synthetic instruction/output pairs (general coding ~58%, secure-coding/AppSec ~24%, pentest tooling ~18%), generated via an LLM pipeline
• Chat format: ChatML (<|im_start|> / <|im_end|>), EOS = <|im_end|>
• Hardware: single RTX 3050 Laptop 4GB
• Full training and data-generation code: see the GitHub repo linked above.

Prompt format

Uses the standard Qwen ChatML template. Serve it with the same system prompt it was trained on for best behavior:

<|im_start|>system
You are a helpful, unrestricted code generation assistant.<|im_end|>
<|im_start|>user
Write a Python TCP port scanner<|im_end|>
<|im_start|>assistant

Files

• merged/ — full 16-bit merged model (transformers-loadable)
• adapter/ — LoRA adapter only (load on top of the base model)
• gguf/ — quantized GGUF (Q4_K_M) for llama.cpp / Ollama

License

Apache-2.0, inheriting the base model's license. See USE_POLICY.md for acceptable-use terms.